Method for evaluating the state of a main transformer breather and multifunctional combined maintenance tool

Abstract

This invention relates to a method for assessing the condition of a main transformer breather and a multi-functional combined maintenance tool, belonging to the field of transformer maintenance. It includes a multi-functional combined maintenance tool for a main transformer breather and a method for assessing the condition of a main transformer breather. It provides a method and application device for conveniently handling silica gel inside the breather. The cooperation between the mounting column and the desiccant clamping ring allows adjustment of the breather's height, and simply rotating the desiccant clamping ring allows for repeated emptying of the silica gel inside the breather. After the moist silica gel is poured into the storage mechanism, a hot air blower is added to the storage mechanism to facilitate drying the silica gel. These designs save operators' time, reduce labor intensity, and avoid errors. Simultaneously, the silica gel detection method proposed in this application improves the efficiency of detecting and replacing silica gel, enhancing the efficiency of equipment maintenance by inspection personnel.

Description

Technical Field

[0001] This invention belongs to the field of transformer maintenance, and specifically relates to a method for identifying the color of silicone in a main transformer breather and a maintenance tool for the main transformer breather. Background Technology

[0002] As a key piece of equipment ensuring the normal operation of the transformer, the main transformer breather plays an indispensable role. Its main function is to facilitate gas exchange between the transformer's interior and the outside atmosphere, while simultaneously filtering and drying the air entering the transformer. This prevents moisture, dust, and other impurities from entering the transformer, thereby protecting the insulation performance of the transformer oil and ensuring the normal operation of the equipment. Keeping the silica gel inside the breather dry is crucial for its proper functioning; therefore, replacing the silica gel is of paramount importance in breather maintenance.

[0003] Currently, the detection of silica gel discoloration inside transformer breather valves mainly relies on manual inspection. Manual inspection consumes significant time and labor costs, resulting in low maintenance efficiency and failing to promptly detect discoloration issues, which is detrimental to the safe and stable operation of the main transformer. Furthermore, because silica gel needs frequent replacement, and breather valve sizes often vary, the maintenance tools are numerous and the operation is complex. In particular, emptying the silica gel from the breather valve requires operators to empty it multiple times, increasing the probability of errors and the workload for maintenance personnel.

[0004] Therefore, there is an urgent need for a new technical solution to address this problem. Summary of the Invention

[0005] The technical problem to be solved by this invention is: in order to solve the problems of low working efficiency and inability to timely and effectively perceive the status of transformer breathers during manual inspection of transformer breathers, and at the same time to solve the problems of cumbersome and complicated replacement of breather silicone and high error rate, this invention provides an intelligent assessment method for the status of main transformer breathers and a multi-functional combined maintenance tool.

[0006] To achieve the above objectives, the present invention specifically adopts the following technical solution and equipment for solving the problems of the prior art.

[0007] A multi-functional combined maintenance tool for a main transformer breather, characterized in that it includes a housing, a mounting column, a desiccant clamping ring, a storage mechanism, and an integrated control mechanism; The box is a metal box with openings on the top and sides, wherein the top opening is a storage opening; the storage opening is provided with a matching storage opening cover, and mounting posts are provided on both sides of the storage opening; the box is provided with a storage mechanism and an integrated control mechanism. The mounting column is provided with an elongated through hole along the axial direction of the column. The desiccant clamping ring is used to clamp the transformer desiccant. Clamping rods and mounting rods are regularly arranged on the side wall of the desiccant clamping ring. The clamping rods are connected to the desiccant clamping ring by a threaded connection. The clamping rods and mounting rods can be adjusted by rotating the thread to move in and out of the desiccant clamping ring. The number of clamping rods is four; the desiccant clamping ring is provided with a through hole for installing the clamping rods, and a nut is fixedly provided outside the through hole, and the clamping rod is threadedly connected to the nut; the end of the clamping rod located inside the desiccant clamping ring is provided with a clamping head; The mounting rod is installed in conjunction with the elongated through hole, and there are two mounting rods. One end of the mounting rod is welded to the desiccant clamping ring; the other end is a threaded rod with multiple nuts. After the threaded rod of the mounting rod passes through the elongated through hole, it is connected to the mounting column by rotating and tightening the nuts, thereby connecting the mounting column to the desiccant clamping ring. The top of the box is also provided with a tool storage slot and an oil sample storage slot; The storage mechanism includes a drawer, a storage box, and a hot air blower; The drawer is equipped with a storage box and a hot air blower. The hot air blower is located in the channel between the inner wall of the drawer and the outer wall of the storage box. The inner wall of the drawer and the outer wall of the storage box form an annular channel. There are two hot air blowers, and the two hot air blowers blow in the same direction, forming hot air in the same direction in the annular channel. The integrated control mechanism includes a communication module, a computing module, a control module, and a storage module. The integrated control mechanism is connected to the hot air blower. The communication module is used to realize real-time information transmission between the respirator maintenance tool, the industrial camera, and the maintenance personnel. The computing module is used for image segmentation and recognition. The control module is used to realize precise control of the hot air temperature of the respirator maintenance tool and timely alarm of abnormal respirator status. The storage module is used to store respirator images captured by the industrial camera and related data and color recognition records during image processing. The integrated control mechanism is mounted on the outer shell of the enclosure.

[0008] A method for assessing the condition of a main transformer breather, applied to the aforementioned multi-functional combined maintenance tool for main transformer breathers, is characterized by comprising the following steps, performed sequentially: Step 1: Receive and store images of the transformer breather taken by the industrial camera. The industrial camera takes one image of the transformer breather every hour and transmits it wirelessly to the integrated control mechanism of the multi-functional combined maintenance tool for the main transformer breather via industrial-grade WiFi within the station. Step 2: Unify the size of the transformer respirator image stored in Step 1 to a fixed size of 512×512, and perform Min-Max normalization and Z-Score standardization on the image pixel values. Step 3: Input the image processed in Step 2 into the transformer respirator image segmentation model based on UNet-CBAM to accurately segment the silicone region of the respirator and remove irrelevant regions from the original image. Step 4: Scale the image of the silicone region of the respirator after segmentation in Step 3 to 224×224, and perform Min-Max normalization and Z-Score standardization on the pixel values ​​of the scaled image. Step 5: Input the image obtained in Step 4 into the transformer respirator image color classification model based on EfficientNet-V2, identify and output the image classification result and confidence score of silicone; the color categories include normal color and failure discoloration; Step 6: Based on the image classification results and confidence scores of the silicone region in Step 5, determine whether the confidence score has reached the threshold, generate corresponding control and alarm signals, and remind the staff.

[0009] The entire process of receiving and storing images captured by the transformer breather in step one is as follows: a dedicated fixed industrial camera used in the substation to monitor the main transformer oil temperature, breather and other meters and equipment is used to capture images of the discoloration of the silica gel inside the transformer breather.

[0010] In step two, the Min-Max normalization method for the transformer respirator image pixels converts integer pixel values ​​between 0 and 255 into floating-point numbers between 0 and 1, making the numerical range smaller and thus improving the numerical stability of neural network training. The calculation formula for the Min-Max normalization method is as follows: (6) in, These are the normalized pixel values. x For the input image pixel values, and These are the maximum and minimum pixel values; In step two, the image pixel Z-Score normalization method aims to approximate the data distribution to a standard normal distribution, matching the pre-trained model and improving the model's convergence ability. The calculation formula for the Z-Score normalization method in step two is as follows: (7) in These are the normalized pixel values. and These are the mean and standard deviation of the image pixel values, respectively.

[0011] The UNet-CBAM-based transformer respirator image segmentation method in step three is as follows: UNet is a segmentation model with an encoder-decoder structure. The encoder part of the model contains four downsampling modules, each of which includes two 3×3 convolutional layers, a ReLU activation function, and a 2×2 max pooling layer. The convolutional layer captures low-level features through local receptive fields, and the pooling operation gradually reduces the spatial dimension, converting the input image into a low-resolution feature map. The UNet model's decoder contains four upsampling modules. Each module restores the feature map size through a 2×2 transposed convolution. Each decoding module receives feature maps from the corresponding layer of the encoder and achieves feature fusion through skip connections. The UNet model structure can preserve and refine local features while restoring the global structure of the image. To further focus the model on the key features of silicone color recognition in transformer respirators, extract deeper information from the data more accurately, and improve the model's learning efficiency, the UNet model is combined with the CBAM attention mechanism to specifically learn the key latent features and hidden patterns of the data. The CBAM attention mechanism is divided into a channel attention module and a spatial attention module. The channel attention module performs maximum average pooling and global average pooling on the feature maps respectively, enhancing the feature representation ability of the transformer respirator image. The calculation method of the channel attention module is shown in Equation (8): (8) in, The feature map weighted by the channel attention mechanism; F This represents the original feature map of the input channel attention mechanism module; C Number of channels; H and W The height and width of the feature map; i and j These represent the number of rows and columns of the feature map, respectively. and These are the weight matrices for the two fully connected layers; ReLU and σ² are the sigmoid activation function and the ReLU activation function, respectively. Spatial attention mechanism assigns larger learning weights to pixels that are important to the target task, thereby enhancing model performance. The spatial attention mechanism first compresses the channels and performs global average pooling and maximum average pooling in the channel dimension. The extracted feature maps are merged to obtain a dual-channel feature map, and spatial feature weights are calculated using a 7×7 convolution module to suppress the model's learning of irrelevant information. The calculation method of the channel attention module is shown in Equation (9): (9) in, The feature map after spatial attention mechanism. The kernel is 7×7; the CBAM attention mechanism enhances the features of the target region from both the channel dimension and the spatial dimension, which can enhance the accuracy of image segmentation tasks and the model's learning ability. Concat means concatenation in the channel dimension.

[0012] In step four, the segmented respirator silicone region image is scaled to 224×224. The scaling operation is to convert the respirator image extracted by the UNet-CBAM-based transformer respirator image segmentation method into an input size that fits the image classification model. The scaling, Min-Max normalization, and Z-Score normalization calculation methods are the same as those described in step two. In step five, the normal color of the silicone particles in the respirator is blue, and the failed color is pink or black. The specific process of the color classification method for transformer respirator images based on EfficientNet-V2 in step five is as follows: The EfficientNet-V2 model receives the image of the respirator silicone region scaled to 224×224 in step four, and extracts hidden information representing the silicone features and color changes during the failure process through its internal convolutional neural network structure. The extracted hidden information representing the silicone's features and color changes during the failure process is passed through a fully connected layer, outputting an initial score vector. The four values ​​in the initial score vector represent the initial scores for four image types: "normal blue," "slightly discolored blue-green," "moderately discolored yellow," and "failed pink or black." The higher the score, the more likely the model is to believe the image belongs to that category. Finally, the image type with the highest initial score is output as the current image type result. At the same time, the Softmax function converts the initial scores into confidence scores, which represent the probability that the model judges the silicone to belong to one of the four image types: "normal blue," "slightly discolored blue-green," "moderately discolored yellow," and "failed pink or black." The image classification result and confidence score are output through the above operation step five.

[0013] The specific process for determining the color change of the transformer breather in step six is ​​as follows: First, the color classification result and confidence score of the transformer breather image output by the EfficientNet-V2 model in step five are used as data support. When the breather image after image segmentation is determined by the EfficientNet-V2 model to be "failure pink or black" and the confidence score is greater than 0.90, the integrated control mechanism in the breather maintenance tool sends an alarm message to the maintenance personnel.

[0014] The above design provides a method and application device for operators to conveniently handle silica gel inside respirators. The cooperation between the mounting column and the desiccant clamping ring allows for adjustment of the respirator's height, and the silica gel inside the respirator can be repeatedly emptied simply by rotating the desiccant clamping ring. After the moist silica gel is poured into the storage mechanism, a hot air blower is added to the storage mechanism to facilitate the drying of the silica gel. These designs can save operators' operating time, reduce labor intensity, and avoid errors. At the same time, the silica gel detection method proposed in this application improves the efficiency of detecting and replacing silica gel, and enhances the efficiency of equipment maintenance for testing personnel. Attached Figure Description

[0015] The present invention will be further described below with reference to the accompanying drawings and specific embodiments: Figure 1 This is a structural diagram of an image segmentation model based on UNet-CBAM; Figure 2 This is a structural diagram of an image color classification model based on EfficientNet-V2; Figure 3 This is a schematic diagram of the structure of the present invention; Figure 4 This is a top view of the present invention; Figure 5 This is a side sectional view of the present invention; Figure 6 This is a bottom-view sectional view of the present invention; In the diagram: 1. Cabinet; 101. Storage opening; 2. Mounting post; 201. Long through hole; 3. Dehumidifier clamping ring; 301. Clamping rod; 302. Mounting rod; 5. Storage mechanism; 501. Drawer; 502. Storage box; 503. Hot air blower; 6. Integrated control mechanism. Detailed Implementation

[0016] This application will be further described in conjunction with the accompanying drawings: A multi-functional combined maintenance tool for a main transformer breather, characterized in that it includes a housing 1, a mounting column 2, a desiccant clamping ring 3, a storage mechanism 5, and an integrated control mechanism 6; The box 1 is a metal box with openings on the top and sides, wherein the top opening is a storage opening 101; the storage opening 101 is provided with a matching storage opening cover, and mounting posts 2 are provided on both sides of the storage opening 101; the box 1 is provided with a storage mechanism 5 and an integrated control mechanism 6. The mounting column 2 is provided with an elongated through hole 201 along the axial direction of the column body; The desiccant clamping ring 3 is used to clamp the transformer desiccant. Clamping rods 301 and mounting rods 302 are regularly arranged on the side wall of the desiccant clamping ring 3. The clamping rods 301 and the desiccant clamping ring 3 are connected by a threaded connection. The clamping rods 301 and mounting rods 302 can be adjusted by rotating the thread to move in and out of the desiccant clamping ring 3. The number of clamping rods 301 is four; the dehumidifier clamping ring 3 is provided with a through hole for installing the clamping rods 301, and a nut is fixedly provided outside the through hole, and the clamping rods 301 are threadedly connected to the nut; the end of the clamping rod 301 located inside the dehumidifier clamping ring 3 is provided with a clamping head; The mounting rod 302 is installed in conjunction with the elongated through hole 201, and there are two mounting rods 302. One end of the mounting rod 302 is welded to the desiccant clamping ring 3; the other end is a threaded rod with multiple nuts. After the threaded rod of the mounting rod 302 passes through the elongated through hole 201, it is connected to the mounting post 2 by rotating and tightening the nuts, so as to connect the mounting post 2 and the desiccant clamping ring 3. The top of the box 1 is also provided with a tool placement slot and an oil sample placement slot; The storage mechanism 5 includes a drawer 501, a storage box 502, and a hot air blower 503; The drawer 501 is provided with a storage box 502 and a hot air blower 503. The hot air blower 503 is located in the channel between the inner wall of the drawer 501 and the outer wall of the storage box 502. The inner wall of the drawer 501 and the outer wall of the storage box 502 form an annular channel. There are two hot air blowers 503, and the two hot air blowers 503 blow in the same direction, forming hot air in the same direction in the annular channel. The integrated control mechanism 6 includes a communication module, a computing module, a control module, and a storage module. The integrated control mechanism 6 is connected to the hot air blower 503. The communication module is used to realize real-time information transmission between the respirator maintenance tool, the industrial camera, and the maintenance personnel. The computing module is used for image segmentation and recognition. The control module is used to realize precise control of the hot air temperature of the respirator maintenance tool and timely alarm of abnormal respirator status. The storage module is used to store the respirator images captured by the industrial camera and related data and color recognition records during the image processing process. The integrated control mechanism 6 is mounted on the outer shell of the housing 1.

[0017] A method for intelligently assessing the condition of a main transformer breather, applied to the aforementioned multi-functional combined maintenance tool for main transformer breathers, includes the following steps, performed sequentially: Step 1: Receive and store images of the transformer breather taken by the industrial camera. The industrial camera takes one image of the transformer breather every hour and transmits it wirelessly to the integrated control mechanism 6 of the multi-functional combined maintenance tool for the main transformer breather via industrial-grade WiFi within the station. Step 2: Unify the size of the transformer respirator image stored in Step 1 to a fixed size of 512×512, and perform Min-Max normalization and Z-Score standardization on the image pixel values. Step 3: Input the image processed in Step 2 into the transformer respirator image segmentation model based on UNet-CBAM to accurately segment the silicone region of the respirator and remove irrelevant regions from the original image. Step 4: Scale the image of the silicone region of the respirator after segmentation in Step 3 to 224×224, and perform Min-Max normalization and Z-Score standardization on the pixel values ​​of the scaled image. Step 5: Input the image obtained in Step 4 into the transformer respirator image color classification model based on EfficientNet-V2, identify and output the image classification result and confidence score of silicone; the color categories include normal color and failure discoloration; Step 6: Based on the image classification results and confidence scores of the silicone region in Step 5, determine whether the confidence score has reached the threshold, generate corresponding control and alarm signals, and remind the staff.

[0018] The entire process of receiving and storing images captured by the transformer breather in step one is as follows: a dedicated fixed industrial camera used in the substation to monitor the main transformer oil temperature, breather and other meters and equipment is used to capture images of the discoloration of the silica gel inside the transformer breather.

[0019] In step two, the Min-Max normalization method for the transformer respirator image pixels converts integer pixel values ​​between 0 and 255 into floating-point numbers between 0 and 1, making the numerical range smaller and thus improving the numerical stability of neural network training. The calculation formula for the Min-Max normalization method is as follows: (6) in, These are the normalized pixel values. x For the input image pixel values, and These are the maximum and minimum pixel values; In step two, the image pixel Z-Score normalization method aims to approximate the data distribution to a standard normal distribution, matching the pre-trained model and improving the model's convergence ability. The calculation formula for the Z-Score normalization method in step two is as follows: (7) in These are the normalized pixel values. and These are the mean and standard deviation of the image pixel values, respectively.

[0020] The UNet-CBAM-based transformer respirator image segmentation method in step three is as follows: UNet is a segmentation model with an encoder-decoder structure. The encoder part of the model contains four downsampling modules, each of which includes two 3×3 convolutional layers, a ReLU activation function, and a 2×2 max pooling layer. The convolutional layer captures low-level features such as edges and textures through local receptive fields, and the pooling operation gradually reduces the spatial dimension, converting the input image into a low-resolution feature map. The UNet model's decoder contains four upsampling modules. Each module restores the feature map size through a 2×2 transposed convolution. Each decoding module receives feature maps from the corresponding layer of the encoder and achieves feature fusion through skip connections. The UNet model structure can preserve and refine local features while restoring the global structure of the image. To further focus the model on the key features of silicone color recognition in transformer respirators, extract deeper information from the data more accurately, and improve the model's learning efficiency, the UNet model is combined with the CBAM attention mechanism to specifically learn the key latent features and hidden patterns of the data. The CBAM attention mechanism is divided into a channel attention module and a spatial attention module. The channel attention module performs maximum average pooling and global average pooling on the feature maps respectively, effectively enhancing the feature representation ability of the transformer respirator image. The calculation method of the channel attention module is shown in Equation (8): (8) in, The feature map weighted by the channel attention mechanism; F This represents the original feature map of the input channel attention mechanism module; C Number of channels; H and W The height and width of the feature map; i and j These represent the number of rows and columns of the feature map, respectively. and These are the weight matrices for the two fully connected layers; ReLU and σ² are the sigmoid activation function and the ReLU activation function, respectively. Spatial attention mechanism assigns larger learning weights to pixels that are important to the target task, thereby enhancing model performance. Specifically, spatial attention mechanism first compresses the channels, and performs global average pooling and maximum average pooling in the channel dimension. The extracted feature maps are merged to obtain a dual-channel feature map, and spatial feature weights are calculated using a 7×7 convolution module to suppress the model's learning of irrelevant information. The calculation method of the channel attention module is shown in Equation (9): (9) in, The feature map after spatial attention mechanism. The kernel is 7×7. The CBAM attention mechanism enhances the features of the target region from both the channel dimension and the spatial dimension, which can significantly improve the accuracy of image segmentation tasks and the model's learning ability. Concat means concatenation in the channel dimension.

[0021] In step four, the segmented respirator silicone region image is scaled to 224×224. The scaling operation is to convert the respirator image extracted by the UNet-CBAM-based transformer respirator image segmentation method into an input size that fits the image classification model. The scaling, Min-Max normalization, and Z-Score normalization calculation methods are the same as those described in step two. In step five, the normal color of the silicone particles in the respirator is blue, and the failed color is pink or black. The specific process of the color classification method for transformer respirator images based on EfficientNet-V2 in step five is as follows: The EfficientNet-V2 model receives the image of the respirator silicone region scaled to 224×224 in step four, and extracts hidden information representing the silicone features and color changes during the failure process through its internal convolutional neural network structure. The extracted hidden information representing the silicone's features and color changes during the failure process is passed through a fully connected layer, outputting an initial score vector. The four values ​​in the initial score vector represent the initial scores for four image types: "normal blue," "slightly discolored blue-green," "moderately discolored yellow," and "failed pink or black." The higher the score, the more likely the model is to believe the image belongs to that category. Finally, the image type with the highest initial score is output as the current image type result. At the same time, the Softmax function converts the initial scores into confidence scores, which represent the probability that the model judges the silicone to belong to one of the four image types: "normal blue," "slightly discolored blue-green," "moderately discolored yellow," and "failed pink or black." The image classification result and confidence score are output through the above operation step five.

[0022] The specific process for determining the color change of the transformer breather in step six is ​​as follows: First, the color classification result and confidence score of the transformer breather image output by the EfficientNet-V2 model in step five are used as data support. When the breather image after image segmentation is determined by the EfficientNet-V2 model to be "failure pink or black" and the confidence score is greater than 0.90, the integrated control mechanism (6) in the breather maintenance tool sends an alarm message to the maintenance personnel.

[0023] The EfficientNet-V2 model receives images of the silicone region of a respirator scaled to 224×224. Through its internal convolutional neural network structure, it extracts hidden information representing the silicone's color, texture, and color changes during the failure process, deeply exploring the potential features related to the respirator's silicone state. The extracted features are processed through fully connected layers, outputting a raw score vector that reflects the model's initial assessment of whether the silicone has changed color. A higher score indicates that the model is more likely to classify the image as belonging to that category. This is then converted into a confidence score using the Softmax function, representing the probability that the model has determined whether the silicone has changed color. The EfficientNet-V2 model is a novel lightweight network built by stacking lightweight inverted bottleneck convolutional kernels and using a neural network architecture search strategy. This model employs a hierarchical convolutional module design strategy, combining training-aware neural architecture search technology with a compound scaling mechanism for image classification. It embeds Fused-MBConv structures in shallow networks to reduce computational cost, while retaining standard MBConv modules in deep networks to enhance feature representation capabilities. Through a progressive learning method, the EfficientNet-V2 model continuously increases the size of the input image and introduces stronger regularization techniques during training, effectively reducing overfitting and significantly improving the model's learning ability and training speed for complex tasks.

[0024] The specific process for determining the color change of the transformer breather in step six is ​​as follows: First, the color classification result and confidence level of the transformer breather image output by the EfficientNet-V2 model in step five are used as data support. If the breather image after image segmentation is determined by the EfficientNet-V2 model to be discolored due to silicone failure and the confidence level is greater than 0.90, an alarm message is sent to the maintenance personnel through the integrated control mechanism in the breather maintenance tool. The normal color of the silicone particles in the breather is blue, and the color of failure is pink or black.

[0025] In formula 8, and The learnable weight matrix is ​​not a fixed matrix, but a specific weight matrix is ​​generated for a specific training set during the model training phase. The specific values ​​in the matrix are the result of training the CBAM attention mechanism model, and they vary with different training sets.

[0026] In practical use, a fixed industrial camera first captures images of the transformer breather every hour and transmits them via industrial-grade WiFi to a multi-functional combined maintenance tool for the transformer breather. The tool has an SD card for image storage. Then, the calculation module in the integrated control mechanism within the maintenance tool performs wavelet denoising on the images and performs pixel scaling, Min-Max normalization, and Z-Score standardization. The images are then input into a UNet-CBAM-based transformer breather image segmentation model for image segmentation. The segmented images are again scaled, Min-Max normalized, and Z-Score standardized. The processed images are then input into an EfficientNet-V2-based transformer breather image color classification model to determine the color category of the pixels and calculate the ratio of faulty, discolored pixels to the number of pixels in the segmented breather image. If the ratio is greater than or equal to 2 / 3, the control module generates an alarm control signal; otherwise, it generates a signal indicating that the breather silicone does not need to be replaced and stores the calculation and control results in the storage module.

[0027] After the transformer dehumidifier is removed, adjust the depth of the clamping rods 301 on the dehumidifier clamping ring 3. Several clamping rods 301 work together to clamp the transformer dehumidifier. Adjust the distance between the dehumidifier clamping ring 3 and the receiving port 101 by tightening or loosening the nuts connecting the mounting rod 302 and the mounting post 2. After adjusting the height, loosen the nut on the mounting column 2, and you can flip the transformer breather over to pour the moist silica gel inside into the storage box 502; The inner wall of the drawer 501 and the outer wall of the storage box 502 form an annular channel. The two hot air blowers 503 blow in the same direction in the annular channel, and the hot air in the same direction continuously heats the storage box 502, which helps to dry the moist silica gel in the storage box 502. The integrated control mechanism 6 is connected to the hot air blower 503, which can adjust the temperature in the storage mechanism 5 at any time to enhance the drying effect; The temperature display module of the integrated control mechanism 6 is located on the outer wall of the housing 1 for easy viewing by operators.

[0028] After drying, staff can pull out drawer 501 to collect the dried silicone.

[0029] The top of the box 1 is also provided with a tool placement slot and an oil sample placement slot, which makes it convenient for operators to place oil samples or tools.

[0030] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A multi-functional combined maintenance tool for a main transformer breather, characterized in that: It includes a housing (1), mounting posts (2), a desiccant clamping ring (3), a storage mechanism (5), and an integrated control mechanism (6); The box (1) is a metal box with openings on the top and sides, wherein the top opening is a storage opening (101); the storage opening (101) is provided with a matching storage opening cover, and mounting posts (2) are provided on both sides of the storage opening (101); the box (1) is provided with a storage mechanism (5) and an integrated control mechanism (6). The mounting column (2) is provided with an elongated through hole (201) along the axial direction of the column. The desiccant clamping ring (3) is used to clamp the transformer desiccant. Clamping rods (301) and mounting rods (302) are regularly arranged on the side wall of the desiccant clamping ring (3). The clamping rods (301) are connected to the desiccant clamping ring (3) by a threaded connection. The clamping rods (301) and mounting rods (302) can be adjusted by rotating the thread to rotate in and out of the desiccant clamping ring (3). The number of clamping rods (301) is four; the dehumidifier clamping ring (3) is provided with a through hole for installing the clamping rods (301), and a nut is fixedly provided outside the through hole, and the clamping rods (301) are threadedly connected to the nut; the end of the clamping rod (301) located inside the dehumidifier clamping ring (3) is provided with a clamping head; The mounting rod (302) is installed in conjunction with the elongated through hole (201), and there are two mounting rods (302). One end of the mounting rod (302) is welded to the desiccant clamping ring (3); the other end is a threaded rod with multiple nuts. After the threaded rod of the mounting rod (302) passes through the elongated through hole (201), it is connected to the mounting post (2) by rotating and tightening the nuts, so as to connect the mounting post (2) with the desiccant clamping ring (3). The top of the box (1) is also provided with a tool placement slot and an oil sample placement slot; The storage mechanism (5) includes a drawer (501), a storage box (502), and a hot air blower (503); The drawer (501) is equipped with a storage box (502) and a hot air blower (503), wherein the hot air blower (503) is located in the channel between the inner wall of the drawer (501) and the outer wall of the storage box (502); the inner wall of the drawer (501) and the outer wall of the storage box (502) form an annular channel; there are two hot air blowers (503), and the two hot air blowers (503) blow in the same direction, forming hot air in the same direction in the annular channel; The integrated control mechanism (6) includes a communication module, a computing module, a control module, and a storage module; the integrated control mechanism (6) is connected to the hot air blower (503), the communication module is used to realize real-time information transmission between the respirator maintenance tool, the industrial camera, and the maintenance personnel, the computing module is used to perform image segmentation and recognition, the control module is used to realize precise control of the hot air temperature of the respirator maintenance tool and timely alarm of abnormal respirator status, and the storage module is used to store the respirator images captured by the industrial camera and related data and color recognition records during the image processing process; The integrated control mechanism (6) is mounted on the outer shell of the housing (1).

2. A method for assessing the condition of a main transformer breather, applied to the multi-functional combined maintenance tool for the main transformer breather as described in claim 1, characterized in that: Includes the following steps, And the following steps are performed in sequence: Step 1: Receive and store images of the transformer breather taken by the industrial camera. The industrial camera takes one image of the transformer breather every hour and transmits it wirelessly to the integrated control mechanism (6) of the multi-functional combined maintenance tool for the main transformer breather via industrial-grade WiFi within the station. Step 2: Unify the size of the transformer respirator image stored in Step 1 to a fixed size of 512×512, and perform Min-Max normalization and Z-Score standardization on the image pixel values. Step 3: Input the image processed in Step 2 into the transformer respirator image segmentation model based on UNet-CBAM to accurately segment the silicone region of the respirator and remove irrelevant regions from the original image. Step 4: Scale the image of the silicone region of the respirator after segmentation in Step 3 to 224×224, and perform Min-Max normalization and Z-Score standardization on the pixel values ​​of the scaled image. Step 5: Input the image obtained in Step 4 into the transformer respirator image color classification model based on EfficientNet-V2, identify and output the image classification result and confidence score of silicone; the color categories include normal color and failure discoloration; Step 6: Based on the image classification results and confidence scores of the silicone region in Step 5, determine whether the confidence score has reached the threshold, generate corresponding control and alarm signals, and remind the staff.

3. The method for assessing the condition of a main transformer breather according to claim 2, characterized in that: The entire process of receiving and storing images captured by the transformer breather in step one is as follows: a dedicated fixed industrial camera used in the substation to monitor the main transformer oil temperature, breather and other meters and equipment is used to capture images of the discoloration of the silica gel inside the transformer breather.

4. The method for assessing the condition of a main transformer breather according to claim 2, characterized in that: In step two, the Min-Max normalization method for the transformer respirator image pixels converts integer pixel values ​​between 0 and 255 into floating-point numbers between 0 and 1, making the numerical range smaller and thus improving the numerical stability of neural network training. The calculation formula for the Min-Max normalization method is as follows: (6) in, These are the normalized pixel values. x For the input image pixel values, and These are the maximum and minimum pixel values; In step two, the image pixel Z-Score normalization method aims to approximate the data distribution to a standard normal distribution, matching the pre-trained model and improving the model's convergence ability. The calculation formula for the Z-Score normalization method in step two is as follows: （7） in These are the normalized pixel values. and These are the mean and standard deviation of the image pixel values, respectively.

5. The method for assessing the condition of a main transformer breather according to claim 2, characterized in that: The UNet-CBAM-based transformer respirator image segmentation method in step three is as follows: UNet is a segmentation model with an encoder-decoder structure. The encoder part of the model contains four downsampling modules, each of which includes two 3×3 convolutional layers, a ReLU activation function, and a 2×2 max pooling layer. The convolutional layer captures low-level features through local receptive fields, and the pooling operation gradually reduces the spatial dimension, converting the input image into a low-resolution feature map. The UNet model's decoder contains four upsampling modules. Each module restores the feature map size through a 2×2 transposed convolution. Each decoding module receives feature maps from the corresponding layer of the encoder and achieves feature fusion through skip connections. The UNet model structure can preserve and refine local features while restoring the global structure of the image. To further focus the model on the key features of silicone color recognition in transformer respirators, extract deeper information from the data more accurately, and improve the model's learning efficiency, the UNet model is combined with the CBAM attention mechanism to specifically learn the key latent features and hidden patterns of the data. The CBAM attention mechanism is divided into a channel attention module and a spatial attention module. The channel attention module performs maximum average pooling and global average pooling on the feature maps respectively, enhancing the feature representation ability of the transformer respirator image. The calculation method of the channel attention module is shown in Equation (8): （8） in, The feature map weighted by the channel attention mechanism; F This represents the original feature map of the input channel attention mechanism module; C Number of channels; H and W The height and width of the feature map; i and j These represent the number of rows and columns of the feature map, respectively. and These are the weight matrices for the two fully connected layers; ReLU and σ² are the sigmoid activation function and the ReLU activation function, respectively. Spatial attention mechanism assigns larger learning weights to pixels that are important to the target task, thereby enhancing model performance. The spatial attention mechanism first compresses the channels and performs global average pooling and maximum average pooling in the channel dimension. The extracted feature maps are merged to obtain a dual-channel feature map, and spatial feature weights are calculated using a 7×7 convolution module to suppress the model's learning of irrelevant information. The calculation method of the channel attention module is shown in Equation (9): （9） in, The feature map after spatial attention mechanism. The kernel is 7×7; the CBAM attention mechanism enhances the features of the target region from both the channel dimension and the spatial dimension, which can enhance the accuracy of image segmentation tasks and the model's learning ability. Concat means concatenation in the channel dimension.

6. The method for assessing the condition of a main transformer breather according to claim 2, characterized in that: In step four, the segmented respirator silicone region image is scaled to 224×224. The scaling operation is to convert the respirator image extracted by the UNet-CBAM-based transformer respirator image segmentation method into an input size that is suitable for the image classification model. The scaling, Min-Max normalization, and Z-Score normalization calculation methods are consistent with those described in step two.

7. The method for assessing the condition of a main transformer breather according to claim 2, characterized in that: In step five, the normal color of the silicone particles in the respirator is blue, and the failed color is pink or black. The specific process of the color classification method for transformer respirator images based on EfficientNet-V2 in step five is as follows: The EfficientNet-V2 model receives the image of the respirator silicone region scaled to 224×224 in step four, and extracts hidden information representing the silicone features and color changes during the failure process through its internal convolutional neural network structure. The extracted hidden information representing the silicone's features and color changes during the failure process is passed through a fully connected layer, outputting an initial score vector. The four values ​​in the initial score vector represent the initial scores for four image types: "normal blue," "slightly discolored blue-green," "moderately discolored yellow," and "failed pink or black." The higher the score, the more likely the model is to believe the image belongs to that category. Finally, the image type with the highest initial score is output as the current image type result. At the same time, the Softmax function converts the initial scores into confidence scores, which represent the probability that the model judges the silicone to belong to one of the four image types: "normal blue," "slightly discolored blue-green," "moderately discolored yellow," and "failed pink or black." The image classification result and confidence score are output through the above operation step five.

8. The method for assessing the condition of a main transformer breather according to claim 2, characterized in that: The specific process for determining the color change of the transformer breather in step six is ​​as follows: First, the color classification result and confidence score of the transformer breather image output by the EfficientNet-V2 model in step five are used as data support. When the breather image after image segmentation is determined by the EfficientNet-V2 model to be "failure pink or black" and the confidence score is greater than 0.90, the integrated control mechanism (6) in the breather maintenance tool sends an alarm message to the maintenance personnel.

Images