A method and system for detecting defects in oil-paper insulation of an oil-immersed transformer

Abstract

The present application relates to the technical field of transformer defect analysis and detection, in particular to an oil-immersed transformer oil-paper insulation defect detection method and system; in view of the problems of poor image quality collected by a submarine in a transformer and low defect identification accuracy, the present application first enhances the insulation defect image through a color weight correction bright channel prior algorithm, then constructs an intelligent defect identification model of an encoder containing differential parallel feature convolution, a skip connection with sparse attention fusion and a decoder of composite splicing of weight calibration channels, and after training by a carbon mark defect data set, detects the images collected by the submarine in real time, solves the problems of uneven image exposure and color distortion, accurately extracts the carbon mark defect features, realizes accurate positioning of irregular edges, improves the defect segmentation accuracy and detection efficiency, and can efficiently monitor the oil-paper insulation state of the transformer, thereby providing protection for the safe operation of the power grid.

Description

Technical Field

[0001] This invention relates to the field of transformer defect analysis and detection technology, specifically to a method and system for detecting defects in the oil-paper insulation of oil-immersed transformers. Background Technology

[0002] Detection of oil-paper insulation defects in oil-immersed transformers is an important part of the pre-commissioning inspection and maintenance of high-voltage equipment. However, because large oil-immersed transformers adopt a sealed shell with internal oil filling to remove air insulation design, the insulation performance of the transformer is usually analyzed indirectly by detecting abnormal electrical quantities. However, this conventional detection method is difficult to accurately and specifically obtain the location and severity of oil-paper insulation defects inside high-voltage equipment. If further methods such as oil drainage and manual drilling are used for inspection, it is not only easy to destroy the original oil-paper insulation environment, which is not conducive to the analysis of the causes of insulation degradation, but it may also lead to long inspection cycles and high engineering costs.

[0003] To address the aforementioned issues, the use of miniature vision robots placed inside the casing of oil-immersed transformers to detect the specific location and severity of oil-paper insulation defects under oil-filled conditions has become an important research direction in the field of internal fault detection. These miniature robots are equipped with signal-shielding ultrasonic and visual sensors, allowing for defect diagnosis without affecting the original oil-paper insulation environment. Using miniature vision robots for oil-paper insulation defect detection not only shortens the inspection cycle and reduces engineering costs but also facilitates the analysis of the causes of oil-paper insulation defects and real-time early warning. However, the image quality acquired by the miniature vision robots during inspection tasks is relatively poor, affecting the accurate identification of transformer oil-paper defects. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for detecting defects in the oil-paper insulation of oil-immersed transformers.

[0005] The technical solution of this invention is as follows: A method for detecting defects in the oil-paper insulation of an oil-immersed transformer includes the following operations: A defect intelligent identification model was constructed. Using the transformer oil paper insulation carbon trace defect dataset, the defect intelligent identification model was trained with weights and the model was exported to obtain a defect intelligent identification training model. The defect intelligent identification training model was then used to predict and process the oil paper insulation images inside the high-voltage equipment collected in real time by the submersible in oil to obtain the oil paper insulation defect detection results. During training, the intelligent defect recognition model processes the following steps: the insulation defect image is enhanced to obtain an enhanced insulation defect image; the enhanced insulation defect image is then processed by an encoder and a decoder to obtain the insulation defect segmentation result; the encoder includes differential parallel feature convolution and max pooling, and the decoder includes convolution, upsampling, and weight calibration channel composite stitching; between corresponding layers of the encoder and decoder, there is a skip connection based on sparse attention fusion, which is used to sparsely attention-fuse the encoded features of the current layer of the encoder with the encoded features of the adjacent layers to obtain the sparse attention-fuse features of the current layer, which are used as the input of the corresponding layer of the decoder.

[0006] The image enhancement process is as follows: using the brightness channel prior theory, the ideal illuminance of each point in the insulation defect image is obtained; according to the order of brightness from large to small, the average value of each color channel is taken from the first percentage of points in the brightness channel of the insulation defect image, and used as the color correction value of each color channel; based on the color correction value of each color channel and the ideal illuminance of each point, the illuminance estimate of each point is obtained; based on the illuminance estimate of each point, the corresponding points in the insulation defect image are corrected to obtain the enhanced insulation defect image.

[0007] During the training phase, the operation of differential parallel feature convolution is as follows: The input is processed by convolution to obtain initial convolutional features. The initial convolutional features are then processed by convolution, angular difference convolution, central difference convolution, vertical difference convolution, and horizontal difference convolution, and then concatenated to obtain differential fusion features. The initial convolutional features are then processed by average pooling, convolution, depthwise separable convolution, and sigmoid activation function to obtain a spatial channel joint attention weight map. The differential fusion features and the spatial channel joint attention weight map are multiplied element-wise, and then added element-wise, convolved, and non-linearly processed with the differential fusion features to obtain parallel convolutional fusion features, which are then used as output.

[0008] In the differential parallel feature convolution operation during the prediction phase, the initial convolution features are weighted and summed with the training weights corresponding to the convolution, angular differential convolution, central differential convolution, vertical differential convolution, and horizontal differential convolution in the defect intelligent recognition training model, respectively, and the result is used as the differential fusion feature.

[0009] The operation of weight calibration channel composite splicing is as follows: the first input is the sparse attention fusion feature of the current level, and the second input is the decoding upsampling feature of the next level; the first input is the main input and the second input is the secondary input, and the second input is the main input and the first input is the secondary input, respectively, and the edge feature space calibration processing is performed to obtain the first edge space calibration feature and the second edge space calibration feature; the first edge space calibration feature and the second edge space calibration feature are spliced ​​to obtain the edge space calibration fusion feature, and the edge feature enhancement is performed to obtain the current level decoding feature.

[0010] During the training phase, the sparse attention fusion operation is as follows: the current level coding features are batch normalized to obtain the current level coding normalized features; the previous level coding features are downsampled to obtain the previous level coding downsampled features; the next level coding features are upsampled to obtain the next level coding upsampled features; the current level coding normalized features, the previous level coding downsampled features, and the next level coding upsampled features of the same scale are added element-wise to obtain the initial coding fusion features; the initial coding fusion features are then averaged at scales of 3×3, 2×2, and 1×1 respectively, and the 1×1 scale is averaged... Pooling is used as the first average pooling result. One result is randomly extracted from the pooling results at the 3×3 and 2×2 scales respectively and adjusted to the 1×1 scale to obtain two 1×1 scale average pooling results, namely the second average pooling result and the third average pooling result. One pooling result is randomly selected from the first average pooling result, the second average pooling result, and the third average pooling result as the sparse attention feature. The sparse attention feature is then processed by dot product with the initial encoding fusion feature to obtain the current level sparse attention fusion feature, which is used to perform weight calibration channel composite splicing with the next level decoding feature.

[0011] In the prediction phase, during the sparse attention fusion operation, after obtaining the initial encoding fusion feature, the initial encoding fusion feature is subjected to global average pooling and then dot product with the initial encoding fusion feature to obtain the final reference attention map, which serves as the sparse attention fusion feature of the current level.

[0012] A system for detecting defects in the oil-paper insulation of an oil-immersed transformer, used to implement the above-mentioned method for detecting defects in the oil-paper insulation of an oil-immersed transformer, includes: The training module is used to construct an intelligent defect recognition model. Using a dataset of carbon trace defects in transformer oil paper insulation, the intelligent defect recognition model is trained with weights and the model is exported to obtain a trained intelligent defect recognition model. During training, the intelligent defect recognition model's processing steps are as follows: the insulation defect image undergoes image enhancement processing to obtain an enhanced insulation defect image; the enhanced insulation defect image is then processed by an encoder and decoder to obtain the insulation defect segmentation result; the encoder includes differential parallel feature convolution and max pooling, and the decoder includes convolution, upsampling, and weight calibration channel composite stitching; between corresponding levels of the encoder and decoder, there is a skip connection based on sparse attention fusion, used to sparsely attention-fuse the encoded features of the current level of the encoder with the encoded features of adjacent levels to obtain the sparse attention-fuse features of the current level, which serve as the input to the corresponding level of the decoder. The prediction module uses a defect intelligent identification training model to perform predictive processing on the oil-paper insulation images inside high-voltage equipment collected in real time by the submersible in oil, and obtains the oil-paper insulation defect detection results.

[0013] An oil-immersed transformer oil-paper insulation defect detection device includes a processor and a memory, wherein the processor executes a computer program stored in the memory to implement the above-mentioned oil-immersed transformer oil-paper insulation defect detection method.

[0014] A computer-readable storage medium for storing a computer program, wherein the computer program, when executed by a processor, implements the above-described method for detecting defects in the oil-immersed transformer's paper insulation.

[0015] The beneficial effects of this invention are as follows: This invention provides a method for detecting defects in the oil-paper insulation of oil-immersed transformers. First, it enhances the insulation defect image using a bright channel prior algorithm with color weight correction. Then, it constructs an intelligent defect recognition model consisting of an encoder with differential parallel feature convolution, a skip connection with sparse attention fusion, and a decoder with weighted channel composite splicing. After training on a carbon trace defect dataset, it detects images acquired in real-time by an underwater vehicle. This method solves the problems of uneven image exposure and color distortion, accurately extracts carbon trace defect features, and achieves precise positioning of irregular edges, improving defect segmentation accuracy and detection efficiency. It can efficiently monitor the oil-paper insulation status of transformers, providing a guarantee for the safe operation of the power grid. Attached Figure Description

[0016] The solutions and advantages of this application will become clear to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of the invention.

[0017] In the attached diagram: Figure 1In this embodiment, the image enhancement effect diagram of the method is shown. In the figure, (a) is the original image and (b) is the enhanced image. Figure 2 This is a schematic diagram illustrating the convolution principles of angular difference convolution, vertical difference convolution, and horizontal difference convolution in the embodiment. Figure 3 As shown in the embodiments, the detection effect diagrams of transformer oil paper insulation defects by the method of this embodiment in scenarios 1 and 2 are respectively. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the exemplary embodiments of this application clearer, the technical solutions in the exemplary embodiments of this application are described clearly and completely below. Obviously, the described exemplary embodiments are only some embodiments of this application, and not all embodiments.

[0019] This embodiment provides a method for detecting defects in the oil-paper insulation of oil-immersed transformers, including the following operations: constructing a defect intelligent identification model; using a transformer oil-paper insulation carbon trace defect dataset to train the defect intelligent identification model with weights and derive the model to obtain a defect intelligent identification training model; using the defect intelligent identification training model to perform predictive processing on the oil-paper insulation images inside high-voltage equipment collected in real time by an oil-immersed submersible to obtain the oil-paper insulation defect detection results, thereby achieving efficient monitoring of the oil-paper insulation status inside high-voltage equipment.

[0020] The defect intelligent identification model consists of an encoder, a decoder, and skip connections.

[0021] During training, the image processing in the intelligent defect recognition model is as follows: the insulation defect image undergoes image enhancement processing to obtain an enhanced insulation defect image; the enhanced insulation defect image is then processed by the encoder and decoder to obtain the insulation defect segmentation result. Between corresponding levels of the encoder and decoder, there is a skip connection based on sparse attention fusion, used to fuse the encoded output of the current level of the encoder with the encoded output of adjacent levels, serving as the input to the corresponding level of the decoder.

[0022] In the prediction process, the image processing in the defect intelligent recognition training model is as follows: the oil-paper insulation image undergoes image enhancement processing to obtain an enhanced insulation defect image; the enhanced insulation defect image is then processed by the encoder and decoder to obtain the oil-paper insulation defect detection result. Between corresponding levels of the encoder and decoder, there is a skip connection based on sparse attention fusion, used to fuse the encoded output of the current level of the encoder with the encoded output of adjacent levels, serving as the input to the corresponding level of the decoder.

[0023] In this embodiment, to enhance the contrast and brightness of carbon trace images and solve the problems of uneven exposure and color distortion, the specific content of the image enhancement processing is designed as follows, taking the training process as an example.

[0024] Step 1: Using the bright channel prior theory, obtain the ideal illumination at each point in the insulation defect image. The calculation formula is as follows: , , , Image of insulation defect. Represents pixel coordinates, subscript Represents the image color channels RGB. Represents pixels y At pixel x Within its neighborhood, For the generated pixels Primary illuminance, , , These are the linear coefficients and bias terms within the local window, respectively. For pixels Ideal illumination.

[0025] , The objective function can be minimized, and its expression is: , To guide the image, This is the regularization parameter.

[0026] Step 2: In order to solve the color shift problem caused by the different absorption and scattering rates of insulating oil to different frequencies of light, according to the order of brightness from large to small, the average value of each color channel is taken from the first percentage (preferably the first 0.1%) of the bright channel in the insulation defect image, and used as the color correction value of each color channel.

[0027] Step 3: Based on the color correction values ​​of each color channel and the ideal illuminance of each point, obtain the illuminance estimate for each point. The calculation formula is as follows: ， ， , , , For pixels xIlluminance estimates for the R, G, and B channels respectively, based on the color correction values. , , These are the color correction values ​​for the R, G, and B channels.

[0028] Step 4: Based on the illuminance estimate for each point, correct the corresponding points in the insulation defect image to obtain the enhanced insulation defect image. The calculation formula is as follows: , Image of insulation defect. Image for illuminance estimation.

[0029] See image enhancement effects Figure 1 After the above image enhancement processing, the overall brightness of the low-light image in oil is significantly improved, the contrast of carbon trace edges is clearer, and the sharpness is significantly improved, making the boundary details of the carbon trace image more prominent. This helps the segmentation model to obtain the edge features of the carbon trace and can meet the requirements of image enhancement under low-light conditions inside high-pressure equipment.

[0030] In this embodiment, the encoder includes differential parallel feature convolution and max pooling. This embodiment enhances the model's ability to express spatial texture features and long-range deep semantic information through differential parallel feature convolution, thereby achieving accurate layer-by-layer extraction of high-frequency information and complex features of carbon trace defects.

[0031] The encoder described above consists of four levels of encoding processing and one feature encoding enhancement processing connected in series. The four levels of encoding processing are as follows: the first level of encoding processing is formed by two differential parallel feature convolutions; the second level of encoding processing is formed by one max pooling and two differential parallel feature convolutions; the third level of encoding processing is formed by one max pooling and two differential parallel feature convolutions; and the fourth level of encoding processing is formed by one max pooling and two differential parallel feature convolutions. The feature encoding enhancement processing is implemented by one max pooling and one differential parallel feature convolution.

[0032] The encoder operates as follows: the input (insulation defect enhancement image) undergoes first-level encoding processing to obtain first-level encoded features; the first-level encoded features undergo second-level encoding processing to obtain second-level encoded features; the second-level encoded features undergo third-level encoding processing to obtain third-level encoded features; the third-level encoded features undergo fourth-level encoding processing to obtain fourth-level encoded features; and the fourth-level encoded features undergo feature encoding enhancement processing to obtain encoded enhanced features.

[0033] During training, the operation of differential parallel feature convolution is as follows: the input is processed by convolution to obtain initial convolutional features; the initial convolutional features are then concatenated after convolution, angular difference convolution, central difference convolution, vertical difference convolution, and horizontal difference convolution to obtain differential fusion features; the differential convolution group can capture finer details by calculating the difference between the input feature map and its feature maps processed by different convolution kernels; simultaneously, the initial convolutional features are processed sequentially by average pooling, convolution, depthwise separable convolution, and sigmoid activation function to obtain spatial channel joint attention. The weight map, while enhancing the representation of spatial details of the target, pays more attention to the contextual interdependencies between distant pixels. This is very beneficial for improving the perception of small targets and occluded targets, thereby improving the detection and localization capabilities of targets with weak features. Finally, the differential fusion feature and the spatial channel joint attention weight map are multiplied element-wise, and then added element-wise, convolved, and non-linearly processed with the differential fusion feature (preferably through the sigmoid activation function) to obtain the parallel convolutional fusion feature, which is used as the output to perform differential parallel feature convolution or max pooling operations.

[0034] Central difference convolution adjusts the weights at the center of the convolution kernel to the difference between the weight value at that point and the sum of the weights at all other positions, thereby highlighting high-frequency areas such as edges and contours. Its expression is: , Input feature map x and output feature map y Current location The output value for the current position. for The neighborhood of, the range of values , For position The weight, The input value is the corresponding position in the neighborhood. This is the input value for the current position.

[0035] See Figure 2 Angular difference convolution obtains the gradient information of the image by subtracting weights arranged in angular order, and then explicitly encodes the gradient prior into the convolutional layer. In addition, horizontal difference convolution and vertical difference convolution calculate the horizontal and vertical gradients by calculating the difference between selected pixel pairs, respectively, enhancing the model's ability to represent and generalize high-frequency information such as image details and texture changes.

[0036] In the differential parallel feature convolution operation during the prediction phase, in order to achieve rich texture feature extraction while meeting the lightweight requirement, the initial convolution features are directly weighted and summed with the training weights corresponding to the convolution, angular difference convolution, central difference convolution, vertical difference convolution, and horizontal difference convolution in the defect intelligent recognition training model, respectively, and the result is used as the differential fusion feature.

[0037] The encoder in this embodiment leverages the high-dimensional spatial resolution of differential convolution groups to enhance feature extraction for complex textures and edges of carbon traces. It also introduces long-distance anchor point attention to more flexibly capture long-distance contextual dependencies and local details, compensating for the limited receptive field of detail-enhanced convolution. Under the premise of meeting computational efficiency requirements, it enables more accurate capture of the features of dendritic defect traces in oil paper insulation, significantly improving the edge segmentation capability of complex carbon trace defects in oil paper insulation.

[0038] In this embodiment, the decoder includes convolution, upsampling, and weight calibration channel composite concatenation. The weight calibration channel composite concatenation can adaptively aggregate boundary information and semantic information to depict more fine-grained carbon trace contours and recalibrate the position of carbon trace edge contours. The entire decoder achieves effective fusion of boundary information and contextual information by enhancing the dynamics and expressiveness of spatial aggregation. It can make up for the missing spatial boundary information in deep semantic features and the missing semantic information in shallow features, and can effectively improve the localization accuracy and segmentation precision of complex carbon trace boundaries in the model.

[0039] Meanwhile, there is a skip connection based on sparse attention fusion between the corresponding layers of the encoder and decoder, which is used to perform sparse attention fusion of the coding features of the current layer of the encoder with the coding features of the adjacent layers to obtain the sparse attention fusion features of the current layer, which are used as the input of the corresponding layer of the decoder.

[0040] The decoder consists of four levels of decoding processing and one convolutional processing in series; the decoding processing is achieved by convolution, upsampling and weight calibration channels. The decoder operates as follows: The encoded enhancement features are convolved and upsampled, then fused with the fourth-level sparse attention feature, and finally concatenated through a weight calibration channel to obtain the fourth-level decoding feature. The fourth-level decoding feature is then convolved and upsampled to obtain the fourth-level decoding upsampled feature, which is then fused with the third-level sparse attention feature and concatenated through a weight calibration channel to obtain the third-level decoding feature. The third-level decoding feature is then convolved and upsampled to obtain the third-level decoding upsampled feature, which is then fused with the second-level sparse attention feature and concatenated through a weight calibration channel to obtain the second-level decoding feature. The second-level decoding feature is then convolved and upsampled to obtain the second-level decoding upsampled feature, which is then fused with the first-level sparse attention feature and concatenated through a weight calibration channel to obtain the first-level decoding feature. Finally, the first-level decoding feature is convolved to obtain the output, which serves as the detection result for the oil-paper insulation defect.

[0041] The operation of weight calibration channel composite splicing is as follows: the first input is the sparse attention fusion feature of the current level, and the second input is the decoding upsampling feature of the next level; the first input is the main input and the second input is the secondary input, and the second input is the main input and the first input is the secondary input, respectively, and the edge feature space calibration processing is performed to obtain the first edge space calibration feature and the second edge space calibration feature; the first edge space calibration feature and the second edge space calibration feature are spliced ​​to obtain the edge space calibration fusion feature, and the edge feature enhancement is performed to obtain the current level decoding feature.

[0042] Taking the first input as the primary input and the second input as the secondary input as an example, the edge feature space calibration process is as follows: The first input is processed by edge feature enhancement and sigmoid activation function to obtain the first edge enhancement nonlinear feature; the first input is multiplied element-wise by the first edge enhancement nonlinear feature, and then added element-wise by the first input to obtain the first branch edge enhancement feature; the second input is processed by edge feature enhancement and sigmoid activation function to obtain the second edge enhancement nonlinear feature; the first input is subtracted element-wise by the first edge enhancement nonlinear feature, and then multiplied element-wise by the second edge enhancement nonlinear feature and the second input to obtain the second branch edge enhancement feature; the second branch edge enhancement feature is added element-wise by the first branch edge enhancement feature and the first input to obtain the first edge space calibration feature.

[0043] Edge feature enhancement can be achieved through convolution, batch normalization, and feature nonlinear processing. Feature nonlinear processing can be implemented using the following formula: , , Represents the nonlinear function on the c-th channel. Location in two-dimensional space Input pixels, Funnel-shaped conditions, Indicates on channel c, with The parameter pooling window is centered, and the area within the yellow box in the image below represents the region. The parameters shared by the window within the same channel act on each pixel and its neighborhood, making the activation output simultaneously dependent on the feature distribution of the local region. Compared to the ReLU activation function commonly used in existing techniques, feature nonlinear processing further enhances the model's focus on the two-dimensional spatial conditions of the image. Based on contextual dependence on each pixel element, it effectively utilizes spatial dependencies when determining the output of the activation function, thereby achieving a more flexible nonlinear response. This solves the problem that the ReLU activation function is limited to the one-dimensional spatial correlation calculation of pixel parameters and is difficult to learn complex spatial information.

[0044] In this embodiment, a skip connection based on sparse attention fusion is provided between corresponding layers of the encoder and decoder. This connection is used to sparsely attentionally fuse the encoded features of the current layer of the encoder with the encoded features of adjacent layers to obtain the sparse attention fused features of the current layer. This fusion of multi-scale input features from adjacent layers enhances the Monte Carlo attention mechanism's ability to capture information at different scales. This cross-scale generated attention map improves the model's ability to learn contextual features, enabling the model to better understand the correlation characteristics between tiny carbon traces and surrounding pixel regions, obtain core features of the carbon trace target at scales independent of the carbon trace target itself, and focus on the precise location of the tiny carbon trace target, thereby improving recognition accuracy. It can better adapt to targets with different sizes and complex shapes in the carbon trace images inside transformers, achieving both contrast enhancement for tiny carbon traces in the segmentation model and compensating for the lack of multi-scale feature extraction capabilities of conventional channel attention mechanisms for carbon traces.

[0045] During the training phase, the sparse attention fusion operation is as follows: The current level encoding features are batch normalized to obtain the current level encoding normalized features; the previous level encoding features are downsampled to obtain the previous level encoding downsampled features; the next level encoding features are upsampled to obtain the next level encoding upsampled features; the current level encoding normalized features, the previous level encoding downsampled features, and the next level encoding upsampled features of the same scale are added element-wise to obtain the initial encoding fusion features; the initial encoding fusion features are then averaged using 3×3, 2×2, and 1×1 scales respectively, with the 1×1 scale average pooling used as the first average pooling result, and then pooled from the 3×3 and 2×2 scales respectively. A result is randomly extracted from the pooling results and adjusted to a 1×1 scale, resulting in two 1×1 scale average pooling results: the second average pooling result and the third average pooling result. For example, a 3×3 scale pooling result contains nine pooling features. One of these nine features is randomly selected and adjusted to a 1×1 scale as the second average pooling result. One of the three average pooling results is then randomly selected as the sparse attention feature. The sparse attention feature is then dot-productted with the initial encoded fusion feature to obtain the current level's sparse attention fusion feature, which is used for weight calibration and channel composite splicing with the next level's decoding features. Multi-scale random pooling is used to generate sparse attention maps, enhancing the contrast of features across multiple levels and improving the transmission capability of minute key features of oil-paper insulation defects. Simultaneously, during the model training and prediction phases, different model structures are adaptively selected, enabling the model to acquire richer defect features during training and significantly reducing computation time during inference, thus facilitating real-time segmentation.

[0046] In the prediction phase, during the sparse attention fusion operation, after obtaining the initial encoding fusion feature, the initial encoding fusion feature is directly multiplied by the initial encoding fusion feature after a fixed 1×1 global average pooling to obtain the final reference attention map, which serves as the sparse attention fusion feature of the current level. By omitting the random selection stage of the multi-scale pooling layer, the model detection speed is improved and the repeatability of the inference results is guaranteed.

[0047] If the current level is the first level, sparse attention fusion does not perform downsampling of the encoded features of the previous level because the first level is the top level. If the current level is the fourth level, sparse attention fusion does not perform upsampling of the encoded features of the next level because the fourth level is the bottom level.

[0048] The results of the oil-immersed transformer oil-paper insulation defect separation and identification in this embodiment can be found in [reference]. Figure 3As can be seen, the method of this embodiment can calibrate the target edge of carbon trace defects in oil paper insulation, improve the contrast of significant carbon trace features, and achieve accurate positioning of irregular edges of carbon trace defects in oil paper insulation.

[0049] This embodiment also provides a system for detecting defects in the oil-paper insulation of oil-immersed transformers, used to implement the above-mentioned method for detecting defects in the oil-paper insulation of oil-immersed transformers, including: The training module is used to construct an intelligent defect recognition model. Using a dataset of carbon trace defects in transformer oil paper insulation, the intelligent defect recognition model is trained with weights and the model is exported to obtain a trained intelligent defect recognition model. During training, the intelligent defect recognition model's processing steps are as follows: the insulation defect image undergoes image enhancement processing to obtain an enhanced insulation defect image; the enhanced insulation defect image is then processed by an encoder and decoder to obtain the insulation defect segmentation result; the encoder includes differential parallel feature convolution and max pooling, and the decoder includes convolution, upsampling, and weight calibration channel composite stitching; between corresponding levels of the encoder and decoder, there is a skip connection based on sparse attention fusion, used to sparsely attention-fuse the encoded features of the current level of the encoder with the encoded features of adjacent levels to obtain the sparse attention-fuse features of the current level, which serve as the input to the corresponding level of the decoder. The prediction module uses a defect intelligent identification training model to perform predictive processing on the oil-paper insulation images inside high-voltage equipment collected in real time by the submersible in oil, and obtains the oil-paper insulation defect detection results.

[0050] This embodiment also provides an oil-immersed transformer oil-paper insulation defect detection device, including a processor and a memory, wherein the processor executes the computer program stored in the memory to implement the above-described oil-immersed transformer oil-paper insulation defect detection method.

[0051] This embodiment also provides a computer-readable storage medium for storing a computer program, wherein the computer program, when executed by a processor, implements the above-described method for detecting defects in the oil-immersed transformer's paper insulation.

[0052] This embodiment provides a method for detecting defects in the oil-paper insulation of oil-immersed transformers. First, it enhances the insulation defect image using a bright channel prior algorithm with color weight correction. Then, it constructs an intelligent defect recognition model consisting of an encoder with differential parallel feature convolution, a skip connection with sparse attention fusion, and a decoder with weighted channel composite splicing. After training on a carbon trace defect dataset, it detects images acquired in real-time by an underwater vehicle. This method solves the problems of uneven image exposure and color distortion, accurately extracts carbon trace defect features, and achieves precise positioning of irregular edges, improving defect segmentation accuracy and detection efficiency. It can efficiently monitor the oil-paper insulation status of transformers, providing a guarantee for the safe operation of the power grid.

[0053] This embodiment provides a method for detecting defects in the oil-paper insulation of an oil-immersed transformer. The image enhancement processing adopts a bright channel prior estimation algorithm with color weight correction, which effectively solves the problems of uneven image exposure and color distortion caused by insulating oil in the dark environment inside the transformer. It significantly improves the brightness and contrast of carbon trace images, highlights boundary details, and lays a high-quality image foundation for subsequent defect feature extraction.

[0054] This embodiment provides a method for detecting defects in the oil-paper insulation of an oil-immersed transformer. The differential parallel feature convolution in the encoder, combined with multi-branch differential convolution and spatial channel joint attention mechanism, accurately extracts high-frequency information and complex features of carbon trace defects. At the same time, it flexibly captures long-distance contextual dependencies and local details, making up for the problem of limited receptive field in traditional convolution and improving the ability to perceive weak features and small target defects.

[0055] This embodiment provides a method for detecting defects in the oil-paper insulation of oil-immersed transformers. Based on sparse attention fusion, the skip connection realizes cross-scale fusion of multi-level features of the encoder, enhances the model's learning ability for multi-scale and complex carbon trace defects, improves the transmission efficiency of small key defect features, and solves the problem of insufficient multi-scale feature extraction capability of conventional channel attention mechanisms.

[0056] This embodiment provides a method for detecting defects in the oil-paper insulation of oil-immersed transformers. The composite splicing of the weight calibration channel of the decoder achieves effective fusion of boundary information and contextual information. It can adaptively calibrate the irregular edges of carbon trace defects, making up for the lack of boundary information in deep semantic feature space and the lack of shallow feature semantic information, and greatly improving the accuracy of defect segmentation and positioning.

[0057] This embodiment provides a method for detecting defects in the oil-paper insulation of oil-immersed transformers. The model adopts adaptive structural designs in the training and inference phases. In the inference phase, calculations are simplified through reparameterization and fixed pooling, achieving lightweight design while ensuring detection accuracy. This meets the online segmentation requirements of real-time image acquisition by underwater vehicles and improves detection efficiency.

[0058] This embodiment provides a method for detecting defects in the oil-paper insulation of oil-immersed transformers. It eliminates the need for draining oil from the transformer or manual drilling for inspection, and achieves visual monitoring of oil-paper insulation defects without damaging the original insulation environment. This shortens the detection cycle, reduces engineering costs and personnel risks, provides reliable technical support for transformer insulation condition assessment and partial discharge diagnosis, and enhances the intelligence and safety of power grid construction and operation.

[0059] While exemplary embodiments of the invention have been described herein, many other variations or modifications conforming to the principles of the invention can be directly determined or derived from the disclosure of this invention without departing from its spirit and scope. Therefore, the scope of the invention should be understood and recognized to cover all such other variations or modifications.

Claims

1. A method for detecting defects in the oil-paper insulation of an oil-immersed transformer, characterized in that, This includes the following operations: A defect intelligent identification model was constructed. Using the transformer oil paper insulation carbon trace defect dataset, the defect intelligent identification model was trained with weights and the model was exported to obtain the defect intelligent identification training model. By using a defect intelligent identification training model, the images of oil-paper insulation inside high-voltage equipment collected in real time by an oil-immersed submersible are predicted and processed to obtain the oil-paper insulation defect detection results. During the training process, the defect intelligent recognition model processes the following steps: the insulation defect image is enhanced to obtain the insulation defect enhanced image; The enhanced image of insulation defects is processed by an encoder and a decoder to obtain the insulation defect segmentation result. The encoder includes differential parallel feature convolution and max pooling, and the decoder includes convolution, upsampling and weight calibration channel composite stitching. There is a skip connection based on sparse attention fusion between the corresponding layers of the encoder and decoder, which is used to sparsely attentionally fuse the encoded features of the current layer of the encoder with the encoded features of the adjacent layers to obtain the sparse attention fusion features of the current layer, which are used as the input of the corresponding layer of the decoder.

2. The method for detecting defects in the oil-paper insulation of an oil-immersed transformer according to claim 1, characterized in that, The image enhancement process is as follows: using the prior theory of the bright channel, the ideal illuminance of each point in the insulation defect image is obtained; according to the order of brightness from large to small, the average value of each color channel is taken from the first percentage of points in the bright channel of the insulation defect image, and used as the color correction value of each color channel; based on the color correction value of each color channel and the ideal illuminance of each point, the illuminance estimate of each point is obtained. Based on the illuminance estimate for each point, the corresponding points in the insulation defect image are corrected to obtain an enhanced insulation defect image.

3. The method for detecting defects in the oil-paper insulation of an oil-immersed transformer according to claim 1, characterized in that, During the training phase, the operation of differential parallel feature convolution is as follows: The input is processed by convolution to obtain initial convolutional features. The initial convolutional features are then concatenated after being processed by convolution, angular difference convolution, central difference convolution, vertical difference convolution, and horizontal difference convolution to obtain differential fusion features. The initial convolutional features are then processed by average pooling, convolution, depthwise separable convolution, and sigmoid activation function to obtain a spatial channel joint attention weight map. The differential fusion features and the spatial channel joint attention weight map are multiplied element-wise, and then added element-wise, convolved, and non-linearly processed with the differential fusion features to obtain parallel convolutional fusion features, which are then used as the output.

4. The method for detecting defects in the oil-paper insulation of an oil-immersed transformer according to claim 3, characterized in that, In the differential parallel feature convolution operation during the prediction phase, the initial convolution features are weighted and summed with the training weights corresponding to the convolution, angular differential convolution, central differential convolution, vertical differential convolution, and horizontal differential convolution in the defect intelligent recognition training model, respectively, and the result is used as the differential fusion feature.

5. The method for detecting defects in the oil-paper insulation of an oil-immersed transformer according to claim 1, characterized in that, The operation of weight calibration channel composite splicing is as follows: The first input is the sparse attention fusion feature of the current layer, and the second input is the decoding upsampling feature of the next layer. Using the first input as the main input and the second input as the secondary input, and using the second input as the main input and the first input as the secondary input, respectively, edge feature space calibration processing is performed to obtain the first edge space calibration feature and the second edge space calibration feature. The first edge space calibration feature and the second edge space calibration feature are concatenated to obtain the edge space calibration fusion feature. After edge feature enhancement, the current level decoding feature is obtained.

6. The method for detecting defects in the oil-paper insulation of an oil-immersed transformer according to claim 1, characterized in that, During the training phase, the operation of sparse attention fusion is as follows: The current level coding features are batch normalized to obtain the current level coding normalized features; The upper-level encoded features are downsampled to obtain the upper-level encoded downsampled features; The next level of encoded features are upsampled to obtain the next level of encoded upsampled features; The initial fusion feature is obtained by adding the current level coding normalized feature, the previous level coding downsampled feature, and the next level coding upsampled feature of the same scale element by element. After the initial encoded fusion features are averaged at scales of 3×3, 2×2, and 1×1, the 1×1 average pooling is taken as the first average pooling result. Then, one result is randomly extracted from the pooling results at scales of 3×3 and 2×2 and adjusted to scale of 1×1 to obtain two average pooling results at scale of 1×1, namely the second average pooling result and the third average pooling result. Choose any one of the following pooling results from the first average pooling result, the second average pooling result, and the third average pooling result as the sparse attention feature; The sparse attention features are multiplied by the initial encoded fusion features to obtain the current level sparse attention fusion features, which are then used to perform weight calibration channel composite splicing with the next level decoding features.

7. The method for detecting defects in the oil-paper insulation of an oil-immersed transformer according to claim 6, characterized in that, In the prediction phase, during the sparse attention fusion operation, after obtaining the initial encoding fusion feature, the initial encoding fusion feature is subjected to global average pooling and then dot product with the initial encoding fusion feature to obtain the final reference attention map, which serves as the sparse attention fusion feature of the current level.

8. A system for detecting defects in the oil-paper insulation of an oil-immersed transformer, used to implement the method for detecting defects in the oil-paper insulation of an oil-immersed transformer as described in claim 1, characterized in that, include: The training module is used to build a defect intelligent identification model. Using the transformer oil paper insulation carbon trace defect dataset, the defect intelligent identification model is trained with weights and the model is exported to obtain the defect intelligent identification training model. During the training process, the defect intelligent recognition model processes the following steps: the insulation defect image is enhanced to obtain the insulation defect enhanced image; The enhanced image of insulation defects is processed by an encoder and a decoder to obtain the insulation defect segmentation result. The encoder includes differential parallel feature convolution and max pooling, and the decoder includes convolution, upsampling and weight calibration channel composite stitching. There is a skip connection based on sparse attention fusion between the corresponding layers of the encoder and decoder, which is used to sparsely attentionally fuse the encoded features of the current layer of the encoder with the encoded features of the adjacent layers to obtain the sparse attention fusion features of the current layer, which are used as the input of the corresponding layer of the decoder. The prediction module uses a defect intelligent identification training model to perform predictive processing on the oil-paper insulation images inside high-voltage equipment collected in real time by the submersible in oil, and obtains the oil-paper insulation defect detection results.

9. A device for detecting defects in the oil-paper insulation of an oil-immersed transformer, characterized in that, It includes a processor and a memory, wherein the processor executes a computer program stored in the memory to implement the oil-immersed transformer oil-paper insulation defect detection method as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, Used to store a computer program, wherein the computer program, when executed by a processor, implements the method for detecting defects in the oil-immersed transformer oil-paper insulation as described in any one of claims 1-7.

Images