A method for identifying oil leakage in power distribution equipment based on deep learning

Abstract

The present invention belongs to the technical field of power distribution equipment, specifically a method for identifying oil leakage in power distribution equipment based on deep learning, including the following specific steps: constructing a data set of images of oil leakage in power distribution equipment; performing feature preprocessing on the data set of images of oil leakage in power distribution equipment by frequency domain enhancement and filtering operations; constructing a model OLM for identifying oil leakage in power distribution equipment based on deep learning, inputting the image of oil leakage in power distribution equipment into the model, and the model automatically identifies and locates the oil leakage position of the power distribution equipment in the image of oil leakage in power distribution equipment. The present invention extracts information from multi-level feature graphs, so that the model can more accurately extract the features of oil leakage in power distribution equipment, thereby effectively improving the accuracy of identifying oil leakage in complex power distribution equipment environments, and through continuous learning and training, the accuracy and efficiency of identifying oil leakage in power distribution equipment can be further improved.

Description

Technical Field

[0001] The present invention relates to the technical field of power distribution equipment, and in particular to a method for identifying oil leakage of power distribution equipment based on deep learning. Background Art

[0002] Distribution equipment is a collection of key equipment in the power system that is responsible for the distribution, control, protection and metering of electric energy, ensuring that electric energy is safely and reliably delivered to various power-consuming equipment. As an indispensable and important part of the power system, the operating status of distribution equipment is directly related to the safety and stability of the entire power grid. However, in actual operation, distribution equipment often leaks oil due to aging, wear or external environmental factors, which not only damages the equipment itself, but may also cause serious consequences such as fire and environmental pollution. Therefore, developing an efficient and accurate method for identifying oil leakage in distribution equipment is of great significance for timely discovering and dealing with oil leakage problems and ensuring the safe operation of the power system.

[0003] The traditional manual inspection method requires inspectors to go to the site in person for inspection, which is not only time-consuming and labor-intensive, but also inefficient. Even if some image acquisition equipment (such as drones, inspection robots, etc.) is used for image acquisition, subsequent image analysis and recognition work still requires manual participation, resulting in low recognition efficiency.

[0004] The existing image processing technology is not ideal for identifying oil leakage in distribution equipment in complex environments. The reason is that there are many types of distribution equipment and complex structures, and the oil leakage phenomenon may be affected by many factors, which limits the ability to accurately identify, classify and locate oil leakage in distribution equipment. Therefore, a method for identifying oil leakage in distribution equipment based on deep learning is invented. This method makes full use of the powerful feature extraction and classification capabilities of deep learning technology. By training the deep learning model, it can automatically learn and identify the characteristics of oil leakage in distribution equipment, realize rapid and accurate identification of oil leakage, and thus ensure the safety and stability of the power system. Summary of the invention

[0005] In view of the above-mentioned and / or existing problems in a method for identifying oil leakage in power distribution equipment based on deep learning, the present invention is proposed.

[0006] Therefore, the purpose of the present invention is to provide a method for identifying oil leakage in distribution equipment based on deep learning, which can solve the above-mentioned existing problems.

[0007] To solve the above technical problems, according to one aspect of the present invention, the present invention provides the following technical solutions:

[0008] A method for identifying oil leakage of power distribution equipment based on deep learning, which includes the following specific steps:

[0009] S1: Construct a dataset of oil leakage images of power distribution equipment;

[0010] S2: frequency domain enhancement and filtering operations are used to perform feature preprocessing on the distribution equipment oil leakage image dataset;

[0011] S3: Construct a distribution equipment oil leakage identification model OLM based on deep learning. Input the distribution equipment oil leakage image into the model, and the model automatically identifies and locates the distribution equipment oil leakage position in the distribution equipment oil leakage image;

[0012] S4: Allocate a distribution equipment oil leakage image dataset, train and verify the distribution equipment oil leakage recognition model;

[0013] S5: Apply the oil leakage identification model of distribution equipment and integrate it into the existing distribution equipment monitoring system for alarm and monitoring.

[0014] As a preferred solution of the method for identifying oil leakage of power distribution equipment based on deep learning described in the present invention, the specific steps of S1 are as follows:

[0015] S11: Use high-definition cameras or inspection equipment such as drones to collect images of oil leakage from distribution equipment, and then build a dataset of oil leakage images from distribution equipment;

[0016] S12: annotating the collected distribution equipment oil leakage image dataset by manual box selection;

[0017] S13: Use the LabelImg image annotation tool to select the oil leakage location in the distribution equipment oil leakage image and label its category as oil leakage.

[0018] As a preferred solution of the method for identifying oil leakage of power distribution equipment based on deep learning described in the present invention, the specific steps of S2 are as follows:

[0019] S21: Based on frequency domain enhancement, the original image is transformed from the spatial domain to the frequency domain using Fourier transform, and the image features are enhanced by modifying the spectral components. After enhancement, the inverse Fourier transform is used to achieve the transformation from the frequency domain to the spatial domain.

[0020] S22: Remove random noise in the image based on filtering operations to improve image quality.

[0021] As a preferred solution of the method for identifying oil leakage of power distribution equipment based on deep learning described in the present invention, the specific steps of S3 are as follows:

[0022] S31: Design a feature perception extraction module FPM module. The FPM module expands the detail feature receptive field of the oil leakage image of the distribution equipment through the feature extraction operation of the residual structure and the GAM global attention mechanism, so as to better capture the oil leakage image information in the image and enhance the model's recognition ability of the oil leakage of the distribution equipment. In addition, the FPM module also increases the randomness of the image features through the shuffle operation, improves the generalization performance of the network, and avoids the gradient explosion problem when the model training weight is updated due to regular data, thereby preventing the final model from overfitting or underfitting;

[0023] S32: Design a dual-branch feature fusion module DBFM module. The DBFM module extracts image features from different angles through dual-branch input, and fuses the oil leakage features of distribution equipment at different scales through superposition operation, so as to improve the model's recognition ability of oil leakage of distribution equipment. Since the dual-branch input structure also considers the redundancy of model recognition, when the feature map input at one end is subject to noise or other interference, the feature map input at the other end can still provide effective information support. Its redundancy makes the model more robust when facing complex and changeable inputs, and can output oil leakage recognition results more stably.

[0024] S33: Design a SP-LOSS function as the loss function of the oil leakage classification locator for distribution equipment. The SP-LOSS function weights the area between the predicted box and the true box, and adds the angle between the predicted box and the true box and the distance between the center point as a penalty term into the loss function calculation, thereby optimizing the possible deviation problem of the loss function.

[0025] S34: Based on a feature perception extraction module FPM module designed by S31, a dual-branch feature fusion module DBFM module designed by S32, and an SP-LOSS function designed by S33, a distribution equipment oil leakage recognition model is proposed. By inputting the distribution equipment oil leakage image, the oil leakage position in the image is automatically identified and located.

[0026] As a preferred solution of the method for identifying oil leakage of power distribution equipment based on deep learning described in the present invention, the specific steps of S31 are as follows:

[0027] S311: Input the distribution equipment oil leakage feature map T1 with a size of H×W×C into a CBL module with a convolution kernel number of C / 2, a size of 3×3, and a padding value of 2 to obtain a distribution equipment oil leakage feature map T2 with a size of H×W×C / 2;

[0028] S312: input T2 into GSConv with a convolution kernel number of C / 2, a size of 3×3, and a padding value of 2, and then input it into the BN layer and the LeakyReLU layer for normalization and nonlinear processing to obtain a distribution equipment oil leakage feature map T3 with a size of H×W×C / 2;

[0029] S313: superimpose the distribution equipment oil leakage feature map T2 with a size of H×W×C / 2 and the distribution equipment oil leakage feature map T3 with a size of H×W×C / 2 along the channel to obtain the distribution equipment oil leakage feature map T4 with a size of H×W×C, then perform a shuffle operation on T4, and then input the result into the GAM global attention mechanism to obtain and output the distribution equipment oil leakage feature map T5 with a size of H×W×C;

[0030] T4=Concat(T2,T3)

[0031] T5=GAM(shuffle(T4))

[0032] Among them, Concat represents the superposition operation along the channel, shuffle represents the shuffle operation, and GAM represents the global attention mechanism.

[0033] As a preferred solution of the method for identifying oil leakage of power distribution equipment based on deep learning described in the present invention, the specific steps of S32 are as follows:

[0034] S321: Input distribution equipment oil leakage feature maps A1 and A2 of the same size of H×W×C from two branches, superimpose A1 and A2 along the channel to obtain distribution equipment oil leakage feature map A3 of size H×W×2C, input A3 into Pw-Conv with 2C convolution kernels, 3×3 size, and padding value of 2, and then input into BN layer and LeakyReLU layer for normalization and nonlinear processing to obtain distribution equipment oil leakage feature map A4 of size H×W×2C;

[0035] A3=Concat(A1,A2)

[0036] A4=LeakyReLU(BN(PwConv(A3)))

[0037] Among them, Concat means the superposition operation along the channel, BN means batch normalization, and LeakyReLU is the activation function;

[0038] S322: The distribution equipment oil leakage feature graphs A1 and A2 of size H×W×C are superimposed with the distribution equipment oil leakage feature graph A4 of size H×W×2C along the channel to obtain the distribution equipment oil leakage feature graph A5 of size H×W×4C, and then A5 is input into Pw-Conv with C convolution kernels, 3×3 size, and padding value 2, and then input into the BN layer and LeakyReLU layer for normalization and nonlinear processing to obtain and output the distribution equipment oil leakage feature graph A6 of size H×W×C;

[0039] A5=Concat(A1,A2,A4)

[0040] A6=LeakyReLU(BN(PwConv(A5)))

[0041] Among them, Concat represents the superposition operation along the channel, BN represents batch normalization, and LeakyReLU is the activation function.

[0042] As a preferred solution of the method for identifying oil leakage in power distribution equipment based on deep learning described in the present invention, wherein: the loss function L in S33 sp-LOSS By L LOSS , L m-LOSS and the geometric loss L s composition;

[0043] The loss function L sp-LOSS The expression is as follows:

[0044] L SP-LOSS =αL LOSS +βL m-LOSS +λ s L s

[0045] Among them, α and β are weight coefficients between 0 and 1, and λ S is the loss weight;

[0046] The L LOSS The function expression is as follows:

[0047]

[0048] Among them, R is the area of ​​the predicted region, R gt The actual area of ​​the region;

[0049] The L m-LOSS The function expression is as follows:

[0050]

[0051] Among them, c is the predicted text box center point, c gtis the center point of the real text box, d 2 (c,c gt ) is the center point c of the predicted box and the center point c of the real box gt The Euclidean distance between the detection box and the real box, p is the diagonal length of the minimum rectangle formed by the intersection of the detection box and the real box;

[0052] The geometric loss L s The function expression is as follows:

[0053]

[0054] in, is the weight of the angle loss and penalty term between the predicted box and the true box, is the angle between the predicted box and the real box;

[0055] The SP-LOSS function can provide a moving direction for the detection frame when the detection frame and the real frame do not overlap; when the detection frame and the real frame overlap, the SP-LOSS function can optimize the distance and angle between the detection frame and the real frame through a penalty term to give an optimal detection frame loss.

[0056] As a preferred solution of the method for identifying oil leakage of power distribution equipment based on deep learning described in the present invention, the specific steps of S34 are as follows:

[0057] S341: Input the oil leakage feature map F1 of the power distribution equipment to the STEM module to obtain the oil leakage feature map F2 of the power distribution equipment, input F2 to the CBL module, and then input it to the FPM module for feature perception extraction to obtain the oil leakage feature map F3 of the power distribution equipment, input F3 to the CBL module and the FPM module to obtain the oil leakage feature map F4 of the power distribution equipment, input F4 to the CBL module and the FPM module to obtain the oil leakage feature map F5 of the power distribution equipment, input F5 to the CBL module and the FPM module to obtain the oil leakage feature map F6 of the power distribution equipment;

[0058] S342: Input the distribution equipment oil leakage feature map F3 into the CBL module to obtain the distribution equipment oil leakage feature map F7, use F5 and F7 as the input of the DBFM module to perform double-branch feature fusion to obtain the distribution equipment oil leakage feature map F8, use F6 and F8 as the input of the DBFM module to obtain the distribution equipment oil leakage feature map F9, use F8 and F9 as the input of the DBFM module to obtain the distribution equipment oil leakage feature map F 10 , F7 and F 10 As the input of DBFM module, the oil leakage characteristic map F of the distribution equipment is obtained. 11 , F3 and F 11 As the input of DBFM module, the oil leakage characteristic map F of the distribution equipment is obtained. 12 ;

[0059] S343: The distribution equipment oil leakage characteristic diagram F9, characteristic diagram F 10 , feature map F 11 and feature map F 12 The oil leakage recognition result is input into the oil leakage classification locator of the distribution equipment, and the designed SP-LOSS function is used as the loss function of the classification locator. Finally, the oil leakage recognition result of the distribution equipment is obtained. The oil leakage recognition result includes the oil leakage category, regression box and confidence of the distribution equipment in the image.

[0060] As a preferred solution of the method for identifying oil leakage of power distribution equipment based on deep learning described in the present invention, the specific steps of S4 are as follows:

[0061] S41: Divide the distribution equipment oil leakage image dataset constructed in S1 into a training set and a validation set in a ratio of 3:1;

[0062] S42: training the distribution equipment oil leakage recognition model constructed in S3, first initializing all neural network parameters and setting model-related training parameters;

[0063] S43: Divide the training set and validation set data into multiple batches, and input one batch of training set data into the algorithm for training each time to obtain the training loss value of the batch;

[0064] S44: Input the validation set into the model in batches to obtain the corresponding batch loss values. The algorithm self-learns and adjusts parameters according to the batch loss values ​​each time until the batch loss values ​​tend to converge. The training of the distribution equipment oil leakage identification model is completed.

[0065] As a preferred solution of the method for identifying oil leakage of power distribution equipment based on deep learning described in the present invention, the specific steps of S5 are as follows:

[0066] S51: After the distribution equipment oil leakage identification model is trained, its model is integrated into the existing distribution equipment monitoring system to ensure that it can receive image or video data from the distribution equipment area in real time;

[0067] S52: Set reasonable alarm thresholds in the system to capture the confidence level of oil leakage events. Once the model identifies that the confidence level of the oil leakage risk exceeds the preset threshold, the system will automatically activate the alarm mechanism, which will quickly record and process key information such as the specific time of the oil leakage and the specific location of the distribution equipment where the leakage occurred. Then, the distribution equipment maintenance and safety assurance personnel will be notified immediately via email and text messages to ensure that the oil leakage problem is discovered in time and effective measures are taken, thereby effectively maintaining the stable operation of the distribution system and the safety of power supply.

[0068] Compared with existing technologies:

[0069] The present invention extracts multi-level feature map information, so that the model can more accurately extract the characteristics of oil leakage in distribution equipment, thereby effectively improving the accuracy of oil leakage identification in complex distribution equipment environments. Through continuous learning and training, the accuracy and efficiency of oil leakage identification in distribution equipment can be further improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] Figure 1 It is a schematic diagram of the process of the present invention;

[0071] Figure 2 This is a network diagram of the oil leakage identification model OLM of the power distribution equipment of the present invention;

[0072] Figure 3 This is a schematic diagram of the FPM module structure of the present invention;

[0073] Figure 4 This is a schematic diagram of the DBFM module structure of the present invention. DETAILED DESCRIPTION

[0074] In order to make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0075] The present invention provides a method for identifying oil leakage in power distribution equipment based on deep learning. Figure 1-Figure 4 , including the following specific steps:

[0076] S1: Construct a dataset of oil leakage images of power distribution equipment;

[0077] The specific steps of S1 are as follows:

[0078] S11: Use high-definition cameras or inspection equipment such as drones to collect images of oil leakage from distribution equipment, and then build a dataset of oil leakage images from distribution equipment;

[0079] S12: annotating the collected distribution equipment oil leakage image dataset by manual box selection;

[0080] S13: Use the LabelImg image annotation tool to select the oil leakage location in the distribution equipment oil leakage image and label its category as oil leakage;

[0081] Among them: labelImg is a tool commonly used for image data annotation, and its annotation method is to select the oil leakage location of the power distribution equipment in the image;

[0082] S2: frequency domain enhancement and filtering operations are used to perform feature preprocessing on the distribution equipment oil leakage image dataset;

[0083] The specific steps of S2 are as follows:

[0084] S21: Based on frequency domain enhancement, the original image is transformed from the spatial domain to the frequency domain using Fourier transform, and the image features are enhanced by modifying the spectral components. After enhancement, the inverse Fourier transform is used to achieve the transformation from the frequency domain to the spatial domain.

[0085] S22: removing random noise in the image based on filtering operation to improve image quality;

[0086] S3: Construct a distribution equipment oil leakage identification model OLM based on deep learning. Input the distribution equipment oil leakage image into the model, and the model automatically identifies and locates the distribution equipment oil leakage position in the distribution equipment oil leakage image;

[0087] Among them: frequency domain enhancement and filtering operations are commonly used feature preprocessing methods;

[0088] Frequency domain enhancement mainly consists of the following three steps:

[0089] 1. Image transformation: First, the original image is transformed from the spatial domain to the frequency domain. This step is achieved through Fourier transform, which can convert the image from the spatial coordinate system to the frequency coordinate system, thereby revealing the spectral characteristics of the image.

[0090] 2. Frequency domain processing: In the frequency domain, various operations are performed on the transform coefficients of the image. These processes can be designed with different filters as needed, such as low-pass filters, high-pass filters, band-pass filters, band-stop filters, etc., as well as more complex filtering algorithms such as homomorphic filtering. The function of these filters is to selectively enhance or suppress the spectrum of the image to achieve the purpose of improving image quality.

[0091] 3. Inverse transform: The image spectrum after frequency domain processing is transformed back to the spatial domain to obtain the enhanced image; this step is achieved through inverse Fourier transform.

[0092] The basic principle of filtering operation is to selectively process the spectrum of the image, that is, to design a suitable filter function so that the transform coefficients of the image are weighted or truncated in the frequency domain, so as to achieve the purpose of image enhancement or denoising.

[0093] The specific steps of S3 are as follows:

[0094] S31: Design a feature perception extraction module FPM module. The FPM module expands the detail feature receptive field of the oil leakage image of the distribution equipment through the feature extraction operation of the residual structure and the GAM global attention mechanism, so as to better capture the oil leakage image information in the image and enhance the model's recognition ability of the oil leakage of the distribution equipment. In addition, the FPM module also increases the randomness of the image features through the shuffle operation, improves the generalization performance of the network, and avoids the gradient explosion problem when updating the model training weights due to regular data, thereby preventing the final model from overfitting or underfitting.

[0095] The specific steps of S31 are as follows:

[0096] S311: Input the distribution equipment oil leakage feature map T1 with a size of H×W×C into a CBL module with a convolution kernel number of C / 2, a size of 3×3, and a padding value of 2 to obtain a distribution equipment oil leakage feature map T2 with a size of H×W×C / 2;

[0097] in:

[0098] The CBL module is a commonly used deep learning module, which consists of a Conv layer (convolution layer), a BN layer (batch normalization layer), and a LeakyReLU (activation function layer). It is mainly used for feature extraction and conversion. It extracts the features of the input data through convolution operations, uses the BN layer for normalization, and provides nonlinear characteristics through the LeakyReLU activation function.

[0099] An image of size H×W×C is an image with a length of H pixels, a width of W pixels, and C channels.

[0100] The number of channels of the feature map obtained after the image is calculated by the CBL module is the same as the number of convolution kernels in CBL;

[0101] S312: input T2 into GSConv with a convolution kernel number of C / 2, a size of 3×3, and a padding value of 2, and then input it into the BN layer and the LeakyReLU layer for normalization and nonlinear processing to obtain a distribution equipment oil leakage feature map T3 with a size of H×W×C / 2;

[0102] Among them: GSConv is a lightweight deep learning convolution module, which is mainly used to reduce the computational complexity and parameter quantity of the model while maintaining or improving the performance of the model;

[0103] S313: superimpose the distribution equipment oil leakage feature map T2 with a size of H×W×C / 2 and the distribution equipment oil leakage feature map T3 with a size of H×W×C / 2 along the channel to obtain the distribution equipment oil leakage feature map T4 with a size of H×W×C, then perform a shuffle operation on T4, and then input the result into the GAM global attention mechanism to obtain and output the distribution equipment oil leakage feature map T5 with a size of H×W×C;

[0104] T4=Concat(T2,T3)

[0105] T5=GAM(shuffle(T4))

[0106] Among them, Concat represents the superposition operation along the channel, shuffle represents the shuffle operation, and GAM represents the global attention mechanism;

[0107] S32: Design a dual-branch feature fusion module DBFM module. The DBFM module extracts image features from different angles through dual-branch input, and fuses the oil leakage features of distribution equipment at different scales through superposition operation, so as to improve the model's recognition ability of oil leakage of distribution equipment. Since the dual-branch input structure also considers the redundancy of model recognition, when the feature map input at one end is subject to noise or other interference, the feature map input at the other end can still provide effective information support. Its redundancy makes the model more robust when facing complex and changeable inputs, and can output oil leakage recognition results more stably.

[0108] The specific steps of S32 are as follows:

[0109] S321: Input distribution equipment oil leakage feature maps A1 and A2 of the same size of H×W×C from two branches, superimpose A1 and A2 along the channel to obtain distribution equipment oil leakage feature map A3 of size H×W×2C, input A3 into Pw-Conv with 2C convolution kernels, 3×3 size, and padding value of 2, and then input into BN layer and LeakyReLU layer for normalization and nonlinear processing to obtain distribution equipment oil leakage feature map A4 of size H×W×2C;

[0110] A3=Concat(A1,A2)

[0111] A4=LeakyReLU(BN(PwConv(A3)))

[0112] Among them, Concat means the superposition operation along the channel, BN means batch normalization, and LeakyReLU is the activation function;

[0113] Among them: Pw-Conv is an existing deep learning convolution module that can reduce the number of model parameters and improve operation efficiency;

[0114] S322: The distribution equipment oil leakage feature maps A1 and A2 of size H×W×C are superimposed with the distribution equipment oil leakage feature map A4 of size H×W×2C along the channel to obtain the distribution equipment oil leakage feature map v5 of size H×W×4C, and then A5 is input into Pw-Conv with C convolution kernels, size 3×3, and padding value 2, and then input into the BN layer and LeakyReLU layer for normalization and nonlinear processing to obtain and output the distribution equipment oil leakage feature map A6 of size H×W×C;

[0115] A5=Concat(A1,A2,A4)

[0116] A6=LeakyReLU(BN(PwConv(A5)))

[0117] Among them, Concat means the superposition operation along the channel, BN means batch normalization, and LeakyReLU is the activation function;

[0118] S33: Design a SP-LOSS function as the loss function of the oil leakage classification locator for distribution equipment. The SP-LOSS function weights the area between the predicted box and the true box, and adds the angle between the predicted box and the true box and the distance between the center point as a penalty term into the loss function calculation, thereby optimizing the possible deviation problem of the loss function.

[0119] Loss function L in S33 sp-LOSS By L LOSS , L m-LOSS and the geometric loss L s composition;

[0120] Loss function L sp-LOSS The expression is as follows:

[0121] L SP-LOSS =αL LOSS +βL m-LOSS +λ s L s

[0122] Among them, α and β are weight coefficients between 0 and 1, and λ s is the loss weight;

[0123] L LOSS The function expression is as follows:

[0124]

[0125] Among them, R is the area of ​​the predicted region, R gt The actual area of ​​the region;

[0126] L m-LOSS The function expression is as follows:

[0127]

[0128] Among them, c is the predicted text box center point, c gt is the center point of the real text box, d 2 (c,c gt ) is the center point c of the predicted box and the center point c of the real box gt The Euclidean distance between the detection box and the real box, p is the diagonal length of the minimum rectangle formed by the intersection of the detection box and the real box;

[0129] Geometric loss L s The function expression is as follows:

[0130]

[0131] in, is the weight of the angle loss and penalty term between the predicted box and the true box, is the angle between the predicted box and the real box;

[0132] The SP-LOSS function can provide a moving direction for the detection frame when the detection frame and the real frame do not overlap; the SP-LOSS function can optimize the distance and angle between the detection frame and the real frame through a penalty term when the detection frame and the real frame overlap, thereby giving an optimal detection frame loss;

[0133] S34: Based on a feature perception extraction module FPM module designed in S31, a dual-branch feature fusion module DBFM module designed in S32, and an SP-LOSS function designed in S33, a distribution equipment oil leakage recognition model is proposed. By inputting the distribution equipment oil leakage image, the oil leakage position in the image is automatically recognized and located;

[0134] The specific steps of S34 are as follows:

[0135] S341: Input the oil leakage feature map F1 of the power distribution equipment to the STEM module to obtain the oil leakage feature map F2 of the power distribution equipment, input F2 to the CBL module, and then input it to the FPM module for feature perception extraction to obtain the oil leakage feature map F3 of the power distribution equipment, input F3 to the CBL module and the FPM module to obtain the oil leakage feature map F4 of the power distribution equipment, input F4 to the CBL module and the FPM module to obtain the oil leakage feature map F5 of the power distribution equipment, input F5 to the CBL module and the FPM module to obtain the oil leakage feature map F6 of the power distribution equipment;

[0136] Among them: STEM module is an existing deep learning module. This module can ensure strong feature expression ability and reduce a large number of parameters;

[0137] S342: Input the distribution equipment oil leakage feature map F3 into the CBL module to obtain the distribution equipment oil leakage feature map F7, use F5 and F7 as the input of the DBFM module to perform double-branch feature fusion to obtain the distribution equipment oil leakage feature map F8, use F6 and F8 as the input of the DBFM module to obtain the distribution equipment oil leakage feature map F9, use F8 and F9 as the input of the DBFM module to obtain the distribution equipment oil leakage feature map F 10 , F7 and F 10 As the input of DBFM module, the oil leakage characteristic map F of the distribution equipment is obtained. 11 , F3 and F 11 As the input of DBFM module, the oil leakage characteristic map F of the distribution equipment is obtained. 12 ;

[0138] S343: The distribution equipment oil leakage characteristic diagram F9, characteristic diagram F 10 , feature map F 11 and feature map F 12 The data is input into the distribution equipment oil leakage classification locator, and the designed SP-LOSS function is used as the loss function of the classification locator. Finally, the distribution equipment oil leakage recognition result is obtained. The oil leakage recognition result includes the distribution equipment oil leakage category, regression box, and confidence of the oil leakage in the image.

[0139] S4: Allocate a distribution equipment oil leakage image dataset, train and verify the distribution equipment oil leakage recognition model;

[0140] The specific steps of S4 are as follows:

[0141] S41: Divide the distribution equipment oil leakage image dataset constructed in S1 into a training set and a validation set in a ratio of 3:1;

[0142] S42: training the distribution equipment oil leakage recognition model constructed in S3, first initializing all neural network parameters and setting model-related training parameters;

[0143] S43: Divide the training set and validation set data into multiple batches, and input one batch of training set data into the algorithm for training each time to obtain the training loss value of the batch;

[0144] S44: Input the validation set into the model in batches to obtain corresponding batch loss values. The algorithm self-learns and adjusts parameters according to the batch loss values ​​each time until the batch loss values ​​tend to converge. The training of the distribution equipment oil leakage identification model is completed.

[0145] S5: Apply the oil leakage identification model of power distribution equipment and integrate it into the existing power distribution equipment monitoring system for alarm and monitoring;

[0146] The specific steps of S5 are as follows:

[0147] S51: After the distribution equipment oil leakage identification model is trained, its model is integrated into the existing distribution equipment monitoring system to ensure that it can receive image or video data from the distribution equipment area in real time;

[0148] S52: Set reasonable alarm thresholds in the system to capture the confidence level of oil leakage events. Once the model identifies that the confidence level of the oil leakage risk exceeds the preset threshold, the system will automatically activate the alarm mechanism, which will quickly record and process key information such as the specific time of the oil leakage and the specific location of the distribution equipment where the leakage occurred. Then, the distribution equipment maintenance and safety assurance personnel will be notified immediately via email and text messages to ensure that the oil leakage problem is discovered in time and effective measures are taken, thereby effectively maintaining the stable operation of the distribution system and the safety of power supply.

[0149] The present invention includes but is not limited to the following embodiments:

[0150] In this embodiment, 564 images of oil leakage of distribution equipment are collected for constructing a data set of oil leakage images of distribution equipment. The LabelImg image annotation tool is used to annotate the coordinate box information of the oil leakage positions in the oil leakage images of the distribution equipment. Each image contains one or more oil leakage position targets, and the corresponding oil leakage targets need to be selected with anchor frames multiple times. After all the oil leakage targets are annotated, 564 annotation files in xml format can be obtained. The name of each annotation file is consistent with the name of the corresponding oil leakage image file. Each annotation file can contain the annotation anchor box information and category information of multiple targets.

[0151] After the labeling is completed, use Python to perform image preprocessing on each distribution equipment oil leakage image in the distribution equipment oil leakage image dataset until all the processing is completed. First, the oil leakage image in the distribution equipment oil leakage image dataset is read in as a standard three-color image image through Python's opencv library, image = cv2.imread (image_path). Then, frequency domain enhancement is performed through a Gaussian filter (this embodiment only takes the Gaussian filter as an example, and other frequency domain enhancement filters are also applicable). image = fft_high_pass_filter (image, 10). Finally, the median filter operation (this embodiment only takes the median filter as an example) is used to process the image and reduce noise. gaussian_blurred = cv2.MedianBlur (image, (5, 5), 0) (the kernel size (5, 5) and standard deviation (0) of the median filter here can also be adjusted to adapt to different noise levels).

[0152] The constructed distribution equipment oil leakage image data set is divided into a training set and a validation set at a ratio of 3:1, that is, 423 images of the training set and 141 images of the distribution equipment oil leakage test set. The distribution equipment oil leakage recognition model designed according to the present invention is used to train the distribution equipment oil leakage images after image preprocessing to complete the construction of the distribution equipment oil leakage recognition model. First, initialize the various parameters and hyperparameters of the algorithm training. This embodiment describes several important parameters in the algorithm training: the root initialized optimizer is the OPT optimizer, the initialization training epoch is 400, the batch_size is 16, and the initial learning rate is 0.002. These parameters need to be adaptively adjusted according to the results of multiple algorithm training until the algorithm modeling achieves the optimal effect.

[0153] After completing the initialization of the basic parameters of the algorithm training, start training the distribution equipment oil leakage identification model. Input the distribution equipment oil leakage image after image preprocessing into the distribution equipment oil leakage identification model of the present invention, assuming that the input is a distribution equipment oil leakage feature map F1. The distribution equipment oil leakage feature map F1 is input into the STEM module to obtain the distribution equipment oil leakage feature map F2. After F2 is input into the CBL module, it is input into the FPM module for feature perception extraction to obtain the distribution equipment oil leakage feature map F3, and F3 is input into the CBL module and the FPM module to obtain the distribution equipment oil leakage feature map F4. F4 is input into the CBL module and the FPM module to obtain the distribution equipment oil leakage feature map F5. F5 is input into the CBL module and the FPM module to obtain the distribution equipment oil leakage feature map F6.

[0154] Input the distribution equipment oil leakage characteristic map F3 into the CBL module to obtain the distribution equipment oil leakage characteristic map F7. Use F5 and F7 as the input of the DBFM module to obtain the distribution equipment oil leakage characteristic map F8. Use F5 and F7 as the input of the DBFM module to obtain the distribution equipment oil leakage characteristic map F8. Use F6 and F8 as the input of the DBFM module to obtain the distribution equipment oil leakage characteristic map F9. Use F8 and F9 as the input of the DBFM module to obtain the distribution equipment oil leakage characteristic map F 10 . 10 As the input of DBFM module, the oil leakage characteristic map F of the distribution equipment is obtained. 11 . 11 As the input of DBFM module, the oil leakage characteristic map F of the distribution equipment is obtained. 12 . The oil leakage characteristic diagram of the distribution equipment F9, F 10 、F 11 、F 12Input into the distribution equipment oil leakage classification locator, use the designed SP-LOSS function as the loss function of the classification locator, the training of the detection algorithm of the present invention is to update the parameters inside the algorithm through the back propagation of the loss function, in the loss function described in the present invention, the target bounding box loss adopts SP-LOSS loss, and the bounding box confidence loss adopts binary cross entropy loss, the above loss function is updated round by round in the gradient back propagation process of algorithm training, the algorithm gradually converges in the round by round training, and the detection accuracy is gradually improved. After all rounds of training are completed, the algorithm will call the divided verification data to verify the algorithm's precision Box_Presice and recall rate Box_Recall, and use mAP50 as the evaluation standard for testing the effectiveness of algorithm training. Finally, the distribution equipment oil leakage recognition result is obtained, and the oil leakage recognition result includes the distribution equipment oil leakage category, regression box, and confidence of the oil leakage in the image.

[0155] After the training of the distribution equipment oil leakage identification model is completed, the model is integrated into the existing distribution equipment monitoring system to ensure that it can receive image or video data from the distribution equipment area in real time. In the system, the alarm threshold is set to 0.7 to capture the confidence of the oil leakage event. Once the model identifies that the confidence of the oil leakage risk exceeds the preset threshold, the system will automatically activate the alarm mechanism. This mechanism will quickly record and process key information such as the specific time of the oil leakage and the specific location of the distribution equipment where the leakage occurred, and then immediately notify the distribution equipment maintenance and safety assurance personnel by email and SMS to ensure that the oil leakage problem is discovered in time and effective measures are taken, thereby effectively maintaining the stable operation of the distribution system and the safety of power supply.

[0156] Although the present invention has been described above with reference to the embodiments, various modifications may be made thereto and parts thereof may be replaced by equivalents without departing from the scope of the present invention. In particular, as long as there is no structural conflict, the various features in the embodiments disclosed in the present invention may be used in combination with each other in any manner, and the fact that these combinations are not exhaustively described in this specification is only for the sake of omitting space and saving resources. Therefore, the present invention is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.

Claims

1. A method for identifying oil leakage in power distribution equipment based on deep learning, characterized in that: The specific steps are as follows: S1: Construct a dataset of oil leakage images of power distribution equipment; S2: frequency domain enhancement and filtering operations are used to perform feature preprocessing on the distribution equipment oil leakage image dataset; S3: Construct a distribution equipment oil leakage identification model OLM based on deep learning. Input the distribution equipment oil leakage image into the model, and the model automatically identifies and locates the distribution equipment oil leakage position in the distribution equipment oil leakage image; S4: Allocate a distribution equipment oil leakage image dataset, train and verify the distribution equipment oil leakage recognition model; S5: Apply the oil leakage identification model of power distribution equipment and integrate it into the existing power distribution equipment monitoring system for alarm and monitoring; The specific steps of S3 are as follows: S31: Design a feature perception extraction module FPM module, which uses the feature extraction operation of the residual structure and the GAM global attention mechanism; S32: Design a dual-branch feature fusion module DBFM module, which extracts image features from different angles through dual-branch input and fuses the oil leakage features of distribution equipment at different scales through superposition operation; S33: Design a SP-LOSS function as the loss function of the oil leakage classification locator for distribution equipment. The SP-LOSS function weights the area between the predicted box and the true box, and adds the angle between the predicted box and the true box and the distance between the center point as a penalty term into the loss function calculation; S34: Based on a feature perception extraction module FPM module designed in S31, a dual-branch feature fusion module DBFM module designed in S32, and an SP-LOSS function designed in S33, a distribution equipment oil leakage recognition model is proposed. By inputting the distribution equipment oil leakage image, the oil leakage position in the image is automatically recognized and located; The specific steps of S31 are as follows: S311: input the oil leakage characteristic graph T1 of the power distribution equipment into the CBL module to obtain the oil leakage characteristic graph T2 of the power distribution equipment; S312: Input T2 into GSConv, and then input it into the BN layer and the LeakyReLU layer for processing to obtain the distribution equipment oil leakage feature map T3; S313: T2 and T3 are superimposed along the channel to obtain a distribution equipment oil leakage feature map T4, and then T4 is shuffled, and the result is input into the GAM global attention mechanism to obtain and output the distribution equipment oil leakage feature map T5; The specific steps of S32 are as follows: S321: Input the distribution equipment oil leakage feature maps A1 and A2 from the two branches, superimpose A1 and A2 along the channel to obtain the distribution equipment oil leakage feature map A3, input A3 into Pw-Conv, and then input into the BN layer and the LeakyReLU layer for processing to obtain the distribution equipment oil leakage feature map A4; S322: A1 and A2 are superimposed with A4 along the channel to obtain the oil leakage feature map A5 of the distribution equipment, and then A5 is input into Pw-Conv, and then input into the BN layer and the LeakyReLU layer for processing to obtain and output the oil leakage feature map A6 of the distribution equipment.

2. According to a method for identifying oil leakage in power distribution equipment based on deep learning in claim 1, it is characterized in that: The specific steps of S1 are as follows: S11: Use high-definition cameras or drone inspection equipment to collect oil leakage images of distribution equipment, and then build a distribution equipment oil leakage image dataset; S12: annotating the collected distribution equipment oil leakage image dataset by manual box selection; S13: Use the LabelImg image annotation tool to select the oil leakage location in the distribution equipment oil leakage image and label its category as oil leakage.

3. According to a method for identifying oil leakage in power distribution equipment based on deep learning in claim 1, it is characterized in that: The specific steps of S2 are as follows: S21: Based on frequency domain enhancement, the original image is transformed from the spatial domain to the frequency domain using Fourier transform, and the image features are enhanced by modifying the spectral components. After enhancement, the inverse Fourier transform is used to achieve the transformation from the frequency domain to the spatial domain. S22: Remove random noise in the image based on filtering operations to improve image quality.

4. The method for identifying oil leakage of power distribution equipment based on deep learning according to claim 1, characterized in that: The loss function L in S33 sp-LOSS By L LOSS , L m-LOSS and the geometric loss L s composition; The loss function L sp-LOSS The expression is as follows: L SP-LOSS =αL LOSS +βL m-LOSS +λ s L s Among them, α and β are weight coefficients between 0 and 1, and λ s is the loss weight; The L LOSS The function expression is as follows: Among them, R is the area of ​​the predicted region, R gt The actual area of ​​the region; The L m-LOSS The function expression is as follows: Among them, c is the predicted text box center point, c gt is the center point of the real text box, d 2 (c,c gt ) is the center point c of the predicted box and the center point c of the real box gt The Euclidean distance between the detection box and the real box, p is the diagonal length of the minimum rectangle formed by the intersection of the detection box and the real box; The geometric loss L s The function expression is as follows: in, is the weight of the angle loss and penalty term between the predicted box and the true box, is the angle between the predicted box and the real box; The SP-LOSS function can provide a moving direction for the detection frame when the detection frame and the real frame do not overlap; the SP-LOSS function can optimize the distance and angle between the detection frame and the real frame through a penalty term when the detection frame and the real frame overlap, thereby giving an optimal detection frame loss.

5. The method for identifying oil leakage of power distribution equipment based on deep learning according to claim 1, characterized in that: The specific steps of S34 are as follows: S341: Input the oil leakage feature map F1 of the power distribution equipment to the STEM module to obtain the oil leakage feature map F2 of the power distribution equipment, input F2 to the CBL module, and then input it to the FPM module for feature perception extraction to obtain the oil leakage feature map F3 of the power distribution equipment, input F3 to the CBL module and the FPM module to obtain the oil leakage feature map F4 of the power distribution equipment, input F4 to the CBL module and the FPM module to obtain the oil leakage feature map F5 of the power distribution equipment, input F5 to the CBL module and the FPM module to obtain the oil leakage feature map F6 of the power distribution equipment; S342: Input the distribution equipment oil leakage feature map F3 into the CBL module to obtain the distribution equipment oil leakage feature map F7, use F5 and F7 as the input of the DBFM module to perform double-branch feature fusion to obtain the distribution equipment oil leakage feature map F8, use F5 and F7 as the input of the DBFM module to obtain the distribution equipment oil leakage feature map F8, use F6 and F8 as the input of the DBFM module to obtain the distribution equipment oil leakage feature map F9, use F8 and F9 as the input of the DBFM module to obtain the distribution equipment oil leakage feature map F 10 , F7 and F 10 As the input of DBFM module, the oil leakage characteristic map F of the distribution equipment is obtained. 11 , F3 and F 11 As the input of DBFM module, the oil leakage characteristic map F of the distribution equipment is obtained. 12 ; S343: The distribution equipment oil leakage characteristic diagram F9, characteristic diagram F 10 , feature map F 11 and feature map F 12 The oil leakage recognition result is input into the oil leakage classification locator of the distribution equipment, and the designed SP-LOSS function is used as the loss function of the classification locator. Finally, the oil leakage recognition result of the distribution equipment is obtained. The oil leakage recognition result includes the oil leakage category, regression box and confidence of the distribution equipment in the image.

6. The method for identifying oil leakage of power distribution equipment based on deep learning according to claim 1, characterized in that: The specific steps of S4 are as follows: S41: Divide the distribution equipment oil leakage image dataset constructed in S1 into a training set and a validation set in a ratio of 3:1; S42: training the distribution equipment oil leakage recognition model constructed in S3, first initializing all neural network parameters and setting model-related training parameters; S43: Divide the training set and validation set data into multiple batches, and input one batch of training set data into the algorithm for training each time to obtain the training loss value of the batch; S44: Input the validation set into the model in batches to obtain the corresponding batch loss values. The algorithm self-learns and adjusts parameters according to the batch loss values ​​each time until the batch loss values ​​tend to converge. The training of the distribution equipment oil leakage identification model is completed.

7. The method for identifying oil leakage of power distribution equipment based on deep learning according to claim 1, characterized in that: The specific steps of S5 are as follows: S51: After the distribution equipment oil leakage identification model is trained, its model is integrated into the existing distribution equipment monitoring system to ensure that it can receive image or video data from the distribution equipment area in real time; S52: Set reasonable alarm thresholds in the system to capture the confidence level of oil leakage events. Once the model identifies that the confidence level of the oil leakage risk exceeds the preset threshold, the system will automatically activate the alarm mechanism, which will quickly record and process key information such as the specific time of the oil leakage and the specific location of the distribution equipment where the leakage occurred. Then, the distribution equipment maintenance and safety assurance personnel will be notified immediately via email and text messages to ensure that the oil leakage problem is discovered in time and effective measures are taken, thereby effectively maintaining the stable operation of the distribution system and the safety of power supply.

Images