Fault diagnosis method and device for transformer, and terminal

Abstract

The invention relates to the technical field of transformer fault detection, in particular to a transformer fault diagnosis method and device and a terminal. The method comprises the steps of obtaining an infrared image of a transformer, performing target detection on the infrared image, positioning a transformer area and cutting the transformer area to obtain a target infrared image, and performing fault recognition on the target infrared image by adopting an improved convolutional neural network to obtain a fault category; wherein in the improved convolutional neural network, an L2 pooling layer is adopted to replace a pooling layer in an original convolutional neural network, and the L2 pooling layer performs square and extraction operation on feature values in a pooling window. According to the invention, the problem of low accuracy and real-time performance caused by fault diagnosis of the transformer in the prior art can be solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of transformer fault detection, and in particular to a transformer fault diagnosis method, device and terminal. BACKGROUND

[0002] With the rapid development of power systems and the continuous improvement of intelligent level, the safe operation of transformers is crucial to ensure the stability of power grids and the reliability of power supply. However, during the long-term operation of transformers, various faults are prone to occur due to environmental factors, load changes, and equipment aging. If these faults are not discovered and addressed in a timely manner, they may lead to performance degradation of the transformer, interruption of system operation, or even cause major safety accidents. Therefore, efficient and accurate fault diagnosis technology has become a core requirement for transformer operation and maintenance management.

[0003] Currently, deep learning technology has shown great potential in image processing and fault diagnosis fields, especially the Convolutional Neural Network (CNN), which is widely used due to its excellent performance in feature extraction and pattern recognition. Therefore, a fault diagnosis method based on machine learning can be used to diagnose faults in transformers. However, traditional CNN models are susceptible to noise interference and loss of feature details when processing infrared images, resulting in low accuracy in fault diagnosis. SUMMARY

[0004] The embodiments of the present application provide a transformer fault diagnosis method, device and terminal to solve the problem of low accuracy and real-time performance caused by fault diagnosis of transformers in the prior art.

[0005] In a first aspect, the embodiments of the present application provide a transformer fault diagnosis method, comprising: obtaining an infrared image of a transformer; performing target detection on the infrared image, locating the transformer region and performing cropping to obtain a target infrared image; adopting an improved convolutional neural network to perform fault recognition on the target infrared image to obtain a fault category; wherein an L2 pooling layer is used to replace the pooling layer in the original convolutional neural network in the improved convolutional neural network, and the L2 pooling layer performs square sum and square root operations on the feature values within the pooling window.

[0006] In a possible implementation, the improved convolutional neural network is used to perform fault recognition on the target infrared image to obtain a fault category, comprising: The target infrared image is input into the improved convolutional neural network, convolution operation is performed on the multiple convolutional layers, multi-dimensional features of the target infrared image are extracted, and a first feature map with the same size as the target infrared image is output; In the Mish activation layer, a Mish function is used to perform nonlinear transformation on features in the first feature map, and a second feature map is output, which has the same size as the first feature map; In the L2 pooling layer, L2 pooling processing is performed on a local input vector of the second feature map, and a third feature map is output, which is a compressed feature map of the second feature map; In the Dropout layer, a preset number of neurons are randomly discarded, and a fourth feature map is output, which has the same size as the third feature map; In the global average pooling layer, the dimensions of the current input feature map are globally pooled, and each channel spatial dimension is reduced to 1, and a feature vector with a spatial dimension of 1*1 is output; In the full connection layer, the feature vector with the spatial dimension of 1*1 is mapped to a preset dimension, realizing the mapping of features to fault categories, and a feature vector with the preset dimension is output; In the Softmax layer, the feature vector with the preset dimension is converted into a probability distribution, and prediction probabilities of various categories are output; In the classification layer, the final fault category is determined according to the prediction probabilities of various categories and is output.

[0007] In a possible implementation, the L2 pooling processing on the local input vector of the second feature map includes: According to The L2 pooling processing is performed on the local input vector of the second feature map; Wherein, represents a feature vector after L2 pooling processing, represents the local input vector, , represents a positive integer, which is the number of local input vectors.

[0008] In a possible implementation, the Mish function is: ; Wherein, represents an output result of the Mish function, represents a feature in the first feature map, represents a hyperbolic tangent function, represents a natural constant.

[0009] In one possible implementation, before performing fault identification on the target infrared image using an improved convolutional neural network to obtain the fault category, the method further includes: The improved convolutional neural network is trained to obtain the trained improved convolutional neural network; during the training process, the learning rate is adaptively adjusted for each parameter using the first-order moment and the second-order moment; bias correction is performed on the first-order moment and the second-order moment respectively; the parameters are updated based on the bias-corrected first-order moment and the bias-corrected second-order moment.

[0010] In one possible implementation, the adaptive adjustment of the learning rate for each parameter using the first and second moments includes: according to The learning rate is adaptively adjusted for each parameter; in, Indicates the first The first moment of the step, Indicates the first The second moment of the step, This represents the exponential decay rate of the first moment. This represents the exponential decay rate of the second moment. Indicates the first The first moment of the step, Indicates the first The second moment of the step, Indicates the first The gradient of the step; Deviation corrections are performed on the first-order moment and the second-order moment, respectively, including: according to Deviation corrections are performed on the first-order moment and the second-order moment respectively; in, This represents the first moment after deviation correction. This represents the second moment after deviation correction; The parameter update based on the first-order moment and the second-order moment after deviation correction includes: according to Proceed to the first Step parameter update; in, This indicates the updated parameters. This indicates the parameters before the update. Indicates the learning rate. This represents the regularization term in the denominator to prevent the denominator from being zero.

[0011] In one possible implementation, target detection is performed on the infrared image to locate the transformer region and crop it to obtain the target infrared image, including: The deep separable convolution technology and the channel shuffle technology are adopted to extract thermal features of the infrared image. According to the thermal features, multi-scale thermal feature fusion is performed through a path aggregation network to generate a multi-resolution feature map. A target center point and a bounding box size in the multi-resolution feature map are predicted to locate a transformer region. The transformer region on the infrared image is cropped to obtain a target infrared image.

[0012] In a possible implementation, after the transformer region on the infrared image is cropped to obtain the target infrared image, the method further includes: The target infrared image is subjected to enhancement processing to obtain an enhanced target infrared image.

[0013] In a second aspect, an embodiment of the present application provides a transformer fault diagnosis apparatus, including: An acquisition module is configured to acquire an infrared image of a transformer. A processing module is configured to perform target detection on the infrared image, locate a transformer region, and crop the transformer region to obtain a target infrared image. A fault recognition module is configured to perform fault recognition on the target infrared image by using an improved convolutional neural network to obtain a fault category, wherein an L2 pooling layer is used to replace a pooling layer in the original convolutional neural network in the improved convolutional neural network, and the L2 pooling layer performs square sum and square root operations on feature values in a pooling window.

[0014] In a third aspect, an embodiment of the present application provides a terminal, including a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor implements steps of the transformer fault diagnosis method according to the first aspect or any possible implementation of the first aspect when executing the computer program.

[0015] In a fourth aspect, an embodiment of the present application provides a computer readable storage medium, which stores a computer program, and the computer program implements steps of the transformer fault diagnosis method according to the first aspect or any possible implementation of the first aspect when executed by a processor.

[0016] The embodiment of the present application provides a transformer fault diagnosis method, device and terminal, the infrared image of the transformer is acquired, then target detection is carried out on the infrared image, the transformer area is located and cropped, the target infrared image is obtained, then the improved convolutional neural network is used for fault identification on the target infrared image, and the fault category is obtained. In the improved convolutional neural network, the L2 pooling layer is used to replace the pooling layer in the original convolutional neural network, the L2 pooling layer carries out square sum and square root operation on the characteristic values in the pooling window, compared with the maximum pooling in the prior art, the L2 pooling integrates the contribution of all elements in the region, and more local feature distribution information is retained, compared with the average pooling, the L2 pooling amplifies the weight of the significant features in the region through square operation, while the correlation of the secondary features is retained, the key details are prevented from being diluted by averaging, thereby the feature detail loss or noise sensitivity problem caused by the traditional pooling method is avoided, the extraction ability of the key fault features of the transformer infrared image is significantly improved, the stability and noise resistance of the feature expression are ensured, and the fault diagnosis demand in a complex environment is adapted. BRIEF DESCRIPTION OF DRAWINGS

[0017] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed to be used in the embodiments or prior art description will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can also be obtained by those skilled in the art without creative labor.

[0018] Figure 1 It is the application scenario diagram of the transformer fault diagnosis method provided by the embodiment of the present application; Figure 2 It is the operation schematic diagram of the target infrared image obtained by the embodiment of the present application; Figure 3 It is the implementation flowchart of the target infrared image obtaining method provided by the embodiment of the present application; Figure 4 It is the schematic diagram of the L2 pooling operation provided by the embodiment of the present application; Figure 5 It is the curve schematic diagram of the Mish function provided by the embodiment of the present application; Figure 6 It is the schematic diagram of the processing of each layer of the L2MCNN model provided by the embodiment of the present application; Figure 7 It is the structural schematic diagram of the transformer fault diagnosis device provided by the embodiment of the present application; Figure 8 It is the structural schematic diagram of the transformer fault diagnosis device provided by another embodiment of the present application; Figure 9 It is the schematic diagram of the terminal provided by the embodiment of the present application. DETAILED DESCRIPTION

[0019] In the following description, for purposes of explanation and not limitation, specific details are set forth such as particular architectures, technologies, techniques, etc. in order to provide a thorough understanding of the embodiments of the present application. However, it will be apparent to those skilled in the art that the present application can be practiced in other embodiments that depart from these specific details. In other instances, detailed descriptions of well-known systems, devices, circuits, and methods are omitted so as not to obscure the description of the present application with unnecessary detail.

[0020] In order to make the objects, technical solutions and advantages of the present application clearer, the following will be described by specific embodiments in conjunction with the accompanying drawings.

[0021] The safe operation of the transformer is crucial to the stability of the power grid and the reliability of power supply. During long-term operation, the transformer is susceptible to various faults due to environmental factors, load changes, and equipment aging. If these faults are not discovered and addressed in a timely manner, they can lead to performance degradation of the equipment, interruption of system operation, and even major safety incidents. Therefore, efficient and accurate fault diagnosis technology has become a core requirement for transformer operation and maintenance management. Currently, fault diagnosis for transformers mainly relies on manual inspection and traditional image processing techniques. However, manual inspection is inefficient and costly, and is limited by the experience and environmental conditions of the inspectors, making it difficult to achieve comprehensive and real-time fault monitoring. Traditional image processing methods, such as threshold segmentation and edge detection based on infrared thermal imaging, can identify fault features to some extent, but they have poor adaptability to low resolution, low contrast, and complex background noise of infrared images, and limited feature extraction capability, resulting in insufficient diagnostic accuracy and robustness In addition, existing machine learning-based fault diagnosis methods often face problems such as insufficient feature extraction, poor model generalization ability, and high computational complexity when processing transformer infrared images, making it difficult to meet the needs of real-time monitoring of edge devices. In recent years, deep learning technology has shown great potential in image processing and fault diagnosis, especially CNN, which has been widely applied due to its excellent performance in feature extraction and pattern recognition.

[0022] However, traditional CNN models are susceptible to noise interference, feature loss, or overfitting when processing infrared images, especially in the context of non-uniform thermal distribution and high dynamic range of transformer infrared images, the limitations of traditional pooling methods and activation functions further constrain the performance of the model. Therefore, there is an urgent need for an efficient and robust fault diagnosis method that optimizes feature extraction and model structure for the characteristics of transformer infrared images to improve the accuracy and real-time performance of fault diagnosis and provide reliable protection for the safe operation of power systems.

[0023] The implementation flowchart of the transformer fault diagnosis method provided by the embodiment of the present application is shown in Figure 1 as follows: Step 101, obtaining an infrared image of a transformer.

[0024] Step 102, target detection is performed on the infrared image, the transformer region is located and cropped, and a target infrared image is obtained.

[0025] In the embodiment, referring to Figure 2 , a NanoDet model is used to perform efficient target detection on the transformer infrared image, which meets the real-time monitoring requirements of edge devices. NanoDet is a lightweight target detection model designed for real-time detection of edge devices.

[0026] In an embodiment, referring to Figure 3 , target detection is performed on the infrared image, the transformer region is located and cropped, and a target infrared image is obtained, which can include the following steps: Step 301, using a depth separable convolution technology and a channel shuffle technology to extract thermal features of the infrared image.

[0027] Optionally, NanoDet uses ShuffleNetV2 as the backbone network, combines depth separable convolution and channel shuffle technology, and significantly reduces the computational complexity. The model parameter amount is only 0.95M, and the weight file is compressed to only 1.8MB. The inference speed on the PC side is more than 100FPS, which is suitable for low-power devices.

[0028] The complexity calculation formula is: ; Among them, represents the total operation amount, , represents the size of the input infrared image, , respectively represent the input channel number and the output channel number, represents the convolution kernel size, represents the group number.

[0029] As can be seen from the above complexity calculation formula, the grouped convolution can significantly reduce the operation amount, which is suitable for the low resolution and low contrast characteristics of the infrared image.

[0030] Step 302, according to the thermal features, multi-scale thermal feature fusion is performed through a path aggregation network to generate a multi-resolution feature map.

[0031] Figure 2In the middle, get the thermal characteristics, input path aggregation network (Path Aggregation Network, PAN), PAN through the bottom-up and top-down path aggregation shallow and deep features, generate multi-resolution feature map, improve the ability of accurate positioning of transformer area. Compared with the commonly used feature pyramid network, NanoDet uses PAN to reduce redundant convolution, only retains efficient path, in order to improve the detection ability of transformer in infrared image.

[0032] Step 303, predict the target center point and boundary box size in the multi-resolution feature map, and locate the transformer region.

[0033] Optionally, the detection head adopts anchor-free design, directly predicts the center point and boundary box size of the target, simplifies the complex post-processing of traditional anchor-based method. Here the target refers to the transformer.

[0034] The detection head adopts generalized focal loss function (Generalized Focal Loss, GFL): ; Among them, Focal loss function value, Indicates the number of samples, Indicates the th real label, Indicates the prediction probability of the th sample, Indicates the modulation parameter.

[0035] Indicates the sample weight, the weight is determined by the error between the model real-time prediction result And the real label , which changes dynamically with the training process. Dynamic adjustment of sample weight, in the early stage of training, the low contrast area of transformer, such as the fuzzy fault edge of thermal distribution in infrared image, the initial prediction error of model Big, loss weight Be automatically enlarged, force the model to focus on learning its features; In the later stage of training, if the prediction accuracy is improved, that is, the error is reduced, the weight is automatically reduced, which avoids overfitting and ensures the accurate positioning of the low contrast area.

[0036] Step 304, crop the transformer region on the infrared image to get the target infrared image.

[0037] In the infrared detection of transformer, NanoDet accurately locates the transformer region, and the output boundary box can be directly used for cropping to provide high-quality sub image for subsequent processing.

[0038] In one embodiment, after cropping the transformer region on the infrared image to obtain the target infrared image, the method further includes: The infrared image of the target is enhanced to obtain the enhanced infrared image of the target.

[0039] To improve the robustness and generalization ability of the subsequent fault diagnosis model, the cropped transformer infrared images need to be processed. Adaptive histogram equalization and bilateral filtering methods are used to enhance the images, and random degree rotation is used to expand the dataset to cope with the complex environment and diverse scenarios of infrared images.

[0040] Step 103: Use an improved convolutional neural network to perform fault identification on the target infrared image and obtain the fault category.

[0041] The improved convolutional neural network replaces the original pooling layer with an L2 pooling layer, which performs square root operations on the feature values ​​within the pooling window.

[0042] The improved convolutional neural network includes multiple convolutional layers, Mish activation layers, L2 pooling layers, Dropout layers, global average pooling layers, and fully connected layers.

[0043] Optionally, the target infrared image here can be an image-enhanced target infrared image.

[0044] In this embodiment, a CNN fault diagnosis model is used for transformer fault diagnosis. CNNs occupy a central position in information and image processing due to their local perception and parameter sharing characteristics. Through convolutional operations, CNNs use multi-scale convolutional kernels to scan the input data layer by layer, efficiently extracting features from low to high levels, capturing the spatial structure and patterns of images, and achieving classification through label mapping. Its typical structure includes an input layer, convolutional layers, activation layers, pooling layers, and fully connected layers, with each layer working collaboratively to adapt to different task requirements.

[0045] The convolutional layer extracts features through convolution operations, and the output formula is: ; in, For a certain layer, the first On the output feature map, the first Line 1 Column elements, and The first The elements of the layer input feature map and their corresponding convolutional kernel weights, It's a bias. For convolution operations, This is used as an activation function to enhance the model's generalization ability. , , , are positive integers.

[0046] The pooling layer is used to compress features and reduce computational cost, usually using max-pooling or average-pooling method, and its specific form is: ; wherein, is the max-pooling operation, is the average-pooling operation, is the element of the k-th layer feature map after pooling, is the size of the pooling region, is a specific element of a feature map in a certain layer of the neural network, wherein the subscript represents the coordinate of the feature map in the spatial dimension, and the superscript represents the channel in the k-th layer.

[0047] After the calculation and processing of the convolutional layer and the pooling layer, the feature map is finally input into the fully connected layer for classification. The fully connected layer maps the extracted features to the fault categories, realizing accurate classification. The hierarchical structure and parameter sharing mechanism of the CNN make it able to efficiently adapt to the complex spatial structure of the transformer infrared image, providing a solid foundation for subsequent improvement.

[0048] However, in the traditional pooling layer, the average pooling weakens the intensity of the fault features by calculating the local region mean, leading to loss of details, and the max-pooling only retains the maximum value of the region, ignoring other potential information, which is easily disturbed by the noise or abnormal value of the infrared image. These limitations weaken the model's ability to perceive complex data when faults occur. Therefore, in order to overcome the defects of traditional pooling, in view of the non-uniformity and complexity of the transformer infrared image, an L2 pooling strategy is proposed in the embodiment. Referring to FIG. 8, it generates a smooth and stable feature representation by calculating the L2 norm of the features in the region. The L2 pooling is defined as the square root of the sum of the squares of the local input vector Figure 4 , is the number of elements in the region, and the formula is: n wherein, represents the feature vector after L2 pooling processing, represents the local input vector, , represents a positive integer, and is the number of local input vectors.

[0049] ​​​​​The L2 pooling operation above squares and square roots the feature values of the local region. Compared with the maximum pooling (only the single maximum value in the region is retained, and the surrounding details are lost), the L2 pooling integrates the contributions of all elements in the region, retaining more distribution information of local features, such as the gradient change of the transformer fault edge and the subtle differences in heat distribution.

[0050] Compared with the average pooling (simple average of elements in the region, easy to blur details), the L2 pooling amplifies the weight of significant features in the region through squaring operation, while retaining the relevance of secondary features, avoiding the dilution of key details by "averaging". For example, significant features such as high temperature signals at fault points.

[0051] There are often temperature fluctuations, device reflections and other noises in infrared images, which are manifested as local small amplitude outliers. L2 pooling suppresses noise in the following ways: The feature values of noise are usually small, and the values are smaller after squaring, with weak contribution to the "square sum", which is further suppressed after square root. While the feature values of fault features, such as thermal anomalies at windings, are larger, the weight is significantly improved after squaring, becoming the dominant factor in the pooling result.

[0052] This "amplifying effective signals and suppressing noise" feature makes the feature expression more stable and reduces the interference of complex environments on fault feature extraction.

[0053] Some fault features of transformers often appear as low contrast and weak signals in infrared images. L2 pooling amplifies the difference between these weak signals and the background through squaring operation, for example, the feature value of the fault point is 2 and the background is 1, the difference is expanded from 1 to 3 after squaring, which enhances the "significance" of key fault features, making it easier for the subsequent network to capture these subtle but important features, thus adapting to the fault diagnosis needs in complex environments.

[0054] In summary, L2 pooling integrates the three mechanisms of "integrating regional information + suppressing noise + amplifying weak features", which not only avoids the loss of details and noise sensitivity of traditional pooling, but also enhances the extraction ability of key fault features in transformer infrared images. It provides more stable and more discriminative feature expression for subsequent fault diagnosis, thereby improving the robustness of feature extraction, and enhancing the noise resistance and stability of feature expression of classification.

[0055] Further, the activation function also affects the performance of the CNN. The Relu or Sigmoid activation function is usually used in the CNN, which preserves the positive value by setting the negative input to 0, and is simple and efficient to calculate, but when the input is negative for a long time, the gradient is always 0, which hinders the parameter update and easily causes the problem of "dead neurons"; the Sigmoid function maps the input to the (0, 1) interval, providing smooth nonlinearity, but when the input value is large, the gradient tends to 0, causing gradient disappearance. These shortcomings make it difficult for the network to train when the infrared image of the transformer has a high dynamic range, and weaken the ability to capture fault features.

[0056] In view of the shortcomings of the traditional activation function, the Mish activation function is introduced in the embodiment, as shown in Figure 5 , which generates efficient and stable feature representation through a dynamic adaptive nonlinear mechanism. The mathematical expression of the Mish function is a combination of the hyperbolic tangent and exponential functions of the input, which is defined as: ; wherein, represents the output result of the Mish function, represents the feature in the first feature map, represents the hyperbolic tangent function, represents the natural constant.

[0057] The Mish activation function combines the linear characteristics of the function and the smooth nonlinear characteristics of the function, forming a continuous function without upper bound, which not only preserves the gradient information of positive and negative values, but also avoids the problem of "dead neurons" of Relu, alleviates the phenomenon of gradient disappearance or explosion, improves the convergence and generalization ability of the model, significantly improves the modeling ability of the model for complex nonlinear structures, and is suitable for mining the complex nonlinear structure of the transformer infrared image dataset.

[0058] In an embodiment, an improved convolutional neural network is used for fault identification of the target infrared image to obtain the fault category, which can include: The target infrared image is input into the improved convolutional neural network, and convolution operation is performed in the multiple convolutional layers to extract the multi-dimensional features of the target infrared image, and a first feature map with the same size as the target infrared image is output. In the Mish activation layer, the Mish function is used to perform nonlinear transformation on the features in the first feature map, and a second feature map is output, which has the same size as the first feature map. In the L2 pooling layer, the local input vectors of the second feature map are subjected to L2 pooling processing, and a third feature map is output, which is a compressed feature map of the second feature map. In the Dropout layer, a preset number of neurons are randomly discarded, and a fourth feature map is output, which has the same size as the third feature map; In the global average pooling layer, the dimensions of the current input feature map are globally pooled, and the spatial dimensions of each channel are reduced to 1, and a feature vector with a spatial dimension of 1*1 is output; In the fully connected layer, the feature vector with a spatial dimension of 1*1 is mapped to a preset dimension, realizing the mapping of features to fault categories, and a feature vector with a preset dimension is output; In the Softmax layer, the feature vector with a preset dimension is converted into a probability distribution, and prediction probabilities of various categories are output. In the classification layer, the final fault category is determined according to the prediction probabilities of various categories and is output.

[0059] Optionally, the local input vector of the second feature map is subjected to L2 pooling processing, including: According to The local input vector of the second feature map is subjected to L2 pooling processing.

[0060] The Mish function is: ; Wherein, represents the output result of the Mish function, represents a feature in the first feature map, represents a hyperbolic tangent function, represents a natural constant.

[0061] In the improved convolutional neural network described above, in order to further enhance the ability of the CNN to extract deep features of the transformer infrared image, an L2-MishCNN (which can be simplified as L2MCNN) model is constructed, that is, a neural network that integrates L2 pooling strategy, Mish activation function, Dropout layer and global average pooling layer. As shown in Figure 6 L2MCNN extracts robust features through multiple layers of convolution and L2 pooling, uses the Mish activation function to enhance the non-linear expression capability, and uses the Dropout layer to avoid over-reliance on a certain feature by discarding neurons. After the feature extraction is completed, the GAP layer is added to directly pool the spatial dimensions of the entire feature map, so that each channel outputs a value. GAP does not use a local window, but directly reduces the spatial dimension to 1, enhancing the robustness of the model to prevent overfitting.

[0062] Optionally, before the improved convolutional neural network is used to identify the fault of the target infrared image and obtain the fault category, it further includes: The improved convolutional neural network is trained to obtain the trained improved convolutional neural network. During the training process, the learning rate is adaptively adjusted for each parameter using the first and second moments. Bias correction is performed on the first and second moments respectively. The parameters are updated based on the bias-corrected first and second moments.

[0063] In this embodiment, the Adam optimizer is used during training to improve the model's accuracy. The core mechanism of the Adam optimizer combines the advantages of momentum and RMSProp, adaptively adjusting the learning rate for each parameter by calculating the first-order moment estimate and the second-order moment estimate. This is particularly relevant in the initial stage. , It is biased towards 0, so bias correction is needed.

[0064] Optionally, the learning rate can be adaptively adjusted for each parameter using the first and second moments, including: according to The learning rate is adaptively adjusted for each parameter; in, Indicates the first The first moment of the step, Indicates the first The second moment of the step, This represents the exponential decay rate of the first moment. This represents the exponential decay rate of the second moment. Indicates the first The first moment of the step, Indicates the first The second moment of the step, Indicates the first The gradient of the step.

[0065] Deviation corrections are applied to the first and second moments, respectively, including: according to Deviation corrections are applied to the first and second moments respectively; in, This represents the first moment after deviation correction. This represents the second moment after deviation correction.

[0066] Based on the first and second moments after deviation correction, the parameters are updated, including: according to Proceed to the first Step parameter update; in, This indicates the updated parameters. This indicates the parameters before the update. Indicates the learning rate. This represents the regularization term in the denominator to prevent the denominator from being zero.

[0067] Next, we describe the output of the fault type in detail using a five-layer improved convolutional neural network.

[0068] If the improved convolutional neural network is five layers, the improved convolutional neural network is used for fault recognition of the target infrared image to obtain the fault category, which can include: The first layer includes three sub-layers: a first convolutional sub-layer, a first Mish activation sub-layer, and a first L2 pooling sub-layer; the operations performed are respectively: In the first convolutional sub-layer, the output target infrared image is subjected to a convolution operation and batch normalization, basic edge and texture features are extracted, and a first feature map is output, the size of the first feature map being the same as that of the target infrared image; for example, if the size of the target infrared image is 224x224x32, the size of the first feature map is 224x224x32.

[0069] In the first Mish activation sub-layer, the first feature map is subjected to a nonlinear transformation through a Mish function, positive and negative gradient information is preserved, feature expression capability is enhanced, a second feature map is output, and the size of the second feature map is the same as that of the first feature map; the size of the second feature map is 224x224x32.

[0070] In the first L2 pooling sub-layer, the second feature map is subjected to an L2 pooling operation, and regional feature amplitudes are integrated to suppress noise, a third feature map is output, and the third feature map is a compressed feature map of the second feature map; for example, the second feature map is subjected to an L2 pooling with a 2x2 window and a step of 2, and the size of the output third feature map is 112x112x32.

[0071] The second layer includes four sub-layers: a second convolutional sub-layer, a second Mish activation sub-layer, a second L2 pooling sub-layer, and a Dropout sub-layer. In the second convolutional sub-layer, the 112x112x32 feature map is further subjected to convolution and batch normalization, and more complex local features such as thermal distribution details are extracted, and the size of the output feature map is 112x112x64.

[0072] In the second Mish activation sub-layer, the nonlinear modeling capability is continuously enhanced through the Mish function, the size of the output feature map remains unchanged, and is 112x112x64. In the second L2 pooling sub-layer, the input feature map is subjected to L2 pooling processing, the feature map is compressed to 56x56x64, and key fault features are strengthened. In the Dropout sublayer, a predetermined number of neurons are randomly discarded to avoid overfitting, and the output size remains 56x56x64. In this embodiment, the predetermined number can be set according to requirements, for example, the predetermined number can be set to 30%, that is, 30% of the neurons are randomly discarded, and it can be understood that the discarded 30% of the neurons are not subjected to fitting processing.

[0073] The third layer includes three sublayers: a third convolutional sublayer, a third Mish activation sublayer, and a third L2 pooling sublayer. In the third convolutional sublayer, the 56x56x64 feature map is subjected to convolution and batch normalization processing to extract deep semantic features, such as the thermal distribution pattern of the transformer winding fault area, and the output feature map size is 56x56x128.

[0074] In the third Mish activation sublayer, the capture of complex nonlinear features is enhanced, and the output feature map size remains unchanged, being 56x56x128. In the third L2 pooling sublayer, L2 pooling processing is performed, and the compressed feature map is output, with a size of 28x28x128, further focusing on the core features of the fault.

[0075] The fourth layer includes three sublayers: a fourth convolutional sublayer, a fourth Mish activation sublayer, and a fourth L2 pooling sublayer. In the fourth convolutional sublayer, the 28x28x128 feature map is subjected to convolution and batch normalization processing to extract high-level abstract features, such as the thermal abnormal area of the transformer, and the output feature map size is 28x28x256. In the fourth Mish activation sublayer, the Mish function is used to model the high dynamic range infrared features, and the output feature map size is 28x28x256. In the fourth L2 pooling sublayer, L2 pooling processing is performed, and the compressed feature map is output, with a size of 14x14x256, refining the fault feature representation.

[0076] The fifth layer includes four sublayers: a global average pooling sublayer, a fully connected sublayer, a Softmax sublayer, and a classification sublayer. In the global average pooling sublayer, the 14x14x256 feature map is subjected to global pooling, reducing the spatial dimension of each channel to 1, and the output feature map size is 1x1x256, enhancing the robustness of the model. In the fully connected sublayer, the 1x1x256 feature is mapped to a 10-dimensional vector, realizing the mapping of the feature to the fault category, and the output feature map size is 1x1x10. In the Softmax sublayer, the 10-dimensional vector is converted into a probability distribution, and the prediction probability of each category is output, and the output feature map size is 1x1x10. In the classification sub-layer, the final fault category is determined according to the probability distribution, and a single classification result is output.

[0077] Through the above hierarchical processing, the infrared image is gradually transformed from the original pixel information into abstract features that can be used for fault diagnosis, and finally the precise fault classification of the transformer fault is realized.

[0078] The following exemplary illustrates the training of the L2MCNN model.

[0079] The transformer infrared image dataset is shown in Table 1, including 2075 infrared images of five types of faults: normal, winding short circuit, overload, ground fault and poor contact of tap switch. The above infrared image dataset is divided into training set and validation set, wherein the training set is used to optimize the parameters of the neural network, and the fault features in the transformer infrared image are effectively fitted through repeated iterations. The validation set is used to evaluate the performance of the model, and the hyperparameters are adjusted according to the evaluation results, so as to ensure the performance stability and generalization ability of the model in the training process. The validation set is completely independent of the training process of the network and does not participate in parameter optimization or hyperparameter adjustment, and can truly reflect the fault recognition rate and generalization ability of the model on unknown data. Therefore, the performance results of the validation set can be used as an objective indicator to measure the performance of the network, and provide a reliable basis for evaluating the effectiveness of the model in actual application scenarios.

[0080] Table 1

[0081] The training batch is set to 32, the training is performed for 150 rounds, the Adam optimization algorithm and the dynamic learning rate adjustment strategy are used to speed up the convergence of the model, the cross-entropy loss function is used to calculate the error, and the initial value of the learning rate is set to 0.001. When the accuracy of the training set reaches 99.5% and the accuracy of the validation set no longer increases in 10 rounds, the training is stopped. The performance of the model is evaluated on the validation set, and the results are used as the final accuracy.

[0082] The L2MCNN model provided in the embodiment will be compared with other fault diagnosis models: the other fault diagnosis models can be six CNNs with a depth of 4 constructed by using different pooling and activation strategies, as shown in Table 2, the L2MCNN model includes an L2 pooling layer and a Mish activation layer, the Method1 model includes an L2 pooling layer and a Relu activation layer, the Method2 model includes an L2 pooling layer and a Sigmoid activation layer, the Method3 model includes a pooling layer and a Mish activation layer, the pooling layer uses a maximum pooling function, the Method4 model includes a pooling layer and a Mish activation layer, the pooling layer uses an average pooling function, the Method5 model includes a pooling layer and a Sigmoid activation layer, the pooling layer uses an average pooling function, and the Method5 model includes a pooling layer and a Relu activation layer, the pooling layer uses a maximum pooling function.

[0083] All methods use the same data set and hardware and software to classify test samples, and the average value of 10 times of diagnosis of the verification set is taken as the final accuracy of transformer fault identification, and it can be known from Table 2 that the accuracy of the L2MCNN model adopted in the embodiment is 98.85%, and the accuracy of the other six methods is from 41.60% to 83.21%.

[0084] Table 2

[0085] The embodiment of the present application provides a transformer fault diagnosis method, which obtains an infrared image of the transformer, then performs target detection on the infrared image, locates the transformer region and performs cropping to obtain a target infrared image, and then uses an improved convolutional neural network to perform fault identification on the target infrared image to obtain a fault category; wherein the L2 pooling layer is used to replace the pooling layer in the original convolutional neural network in the improved convolutional neural network, the L2 pooling layer performs square sum and square root operation on the feature values in the pooling window, avoids the feature detail loss or noise sensitivity problem caused by the traditional pooling method, significantly improves the extraction ability of the key fault features of the transformer infrared image, ensures the stability and noise resistance of the feature expression, and adapts to the fault diagnosis demand in complex environment.

[0086] In the embodiment of the present application, the Mish activation function is used, which combines the characteristics of linearity and nonlinearity, alleviates the gradient disappearance and “dead neuron” problem of the traditional activation function, enhances the modeling ability of the improved convolutional neural network model to the complex nonlinear structure of the infrared image, improves the convergence speed and the accuracy of fault feature capture. The L2MCNN model fuses L2 pooling, Mish activation, Dropout and GAP, and optimizes the training process through the Adam optimizer and the dynamic learning rate adjustment strategy, and the result shows excellent generalization ability and practical application value.

[0087] In the embodiment of the present application, when acquiring the target infrared image, the NanoDet model is adopted, and the infrared image thermal imaging characteristics are adapted by combining the ShuffleNetV2 and PAN network, the transformer region is quickly and accurately located, the calculation complexity and time cost of subsequent processing are significantly reduced, and technical support is provided for real-time monitoring. Further, the target infrared image is preprocessed through adaptive histogram equalization, bilateral filtering and random rotation, the image quality and data set diversity are effectively improved, and the adaptability of the model to complex scenes is further enhanced, laying a foundation for high-precision fault diagnosis.

[0088] It should be understood that the size of the serial number of each step in the above embodiment does not mean the order of execution, and the execution order of each process should be determined according to its function and internal logic, and should not constitute any limitation on the implementation process of the embodiment of the present application.

[0089] The following is a device embodiment of the present application. For details not described in detail, please refer to the corresponding method embodiments described above.

[0090] Figure 7 A structure diagram of a transformer fault diagnosis device provided by the embodiment of the present application is shown. For ease of illustration, only the parts related to the embodiment of the present application are shown, and the details are as follows: As shown in Figure 7 The transformer fault diagnosis device 7 includes an acquisition module 71, a processing module 72 and a fault identification module 73.

[0091] The acquisition module 71 is configured to acquire an infrared image of the transformer. The processing module 72 is configured to perform target detection on the infrared image, locate the transformer region and crop it to obtain a target infrared image. The fault identification module 73 is configured to use an improved convolutional neural network to identify the fault of the target infrared image to obtain a fault category; wherein the L2 pooling layer is used to replace the pooling layer in the original convolutional neural network in the improved convolutional neural network, and the L2 pooling layer performs square sum and square root operation on the feature values in the pooling window.

[0092] In one possible implementation, when the processing module 72 uses the improved convolutional neural network to identify the fault of the target infrared image to obtain the fault category, it is configured to: input the target infrared image into the improved convolutional neural network, perform convolution operation in multiple convolution layers, extract multi-dimensional features of the target infrared image, and output a first feature map with the same size as the target infrared image; In the Mish activation layer, the Mish function is used to perform nonlinear transformation on the features in the first feature map, and output a second feature map, which has the same size as the first feature map. In the L2 pooling layer, the local input vector of the second feature map is subjected to L2 pooling processing, and a third feature map is output, the third feature map being a compressed feature map of the second feature map; In the Dropout layer, a preset number of neurons are randomly discarded, and a fourth feature map is output, the fourth feature map having the same size as the third feature map; In the global average pooling layer, the dimensions of the current input feature map are subjected to global pooling, and each channel spatial dimension is reduced to 1, and a feature vector with a spatial dimension of 1*1 is output; In the full connection layer, the feature vector with a spatial dimension of 1*1 is mapped to a preset dimension, realizing mapping of the feature to a fault category, and a feature vector of the preset dimension is output; In the Softmax layer, the feature vector of the preset dimension is converted into a probability distribution, and prediction probabilities of various categories are output; In the classification layer, a final fault category is determined according to the prediction probabilities of various categories and is output.

[0093] In a possible implementation, when the processing module 72 subjects the local input vector of the second feature map to L2 pooling processing, the processing module 72 is configured to: According to subject the local input vector of the second feature map to L2 pooling processing; wherein, denotes the feature vector after L2 pooling processing, denotes the local input vector, , denotes a positive integer, and is the number of local input vectors.

[0094] In a possible implementation, the Mish function is: ; wherein, denotes an output result of the Mish function, denotes a feature in the first feature map, denotes a hyperbolic tangent function, denotes a natural constant.

[0095] In a possible implementation, referring to Figure 8 , before the fault identification module 73 adopts the improved convolutional neural network to identify a fault of the target infrared image and obtains a fault category, the method further includes: a model training module 74; The model training module 74 is used to train the improved convolutional neural network to obtain the trained improved convolutional neural network. During the training of the improved convolutional neural network, the learning rate is adaptively adjusted for each parameter using the first and second moments. Bias correction is performed on the first and second moments respectively. The parameters are updated based on the bias-corrected first and second moments.

[0096] In one possible implementation, when the model training module 74 adaptively adjusts the learning rate for each parameter using the first and second moments, it is used for: according to The learning rate is adaptively adjusted for each parameter; in, Indicates the first The first moment of the step, Indicates the first The second moment of the step, This represents the exponential decay rate of the first moment. This represents the exponential decay rate of the second moment. Indicates the first The first moment of the step, Indicates the first The second moment of the step, Indicates the first The gradient of the step; When model training module 74 performs bias correction on the first and second moments respectively, it is used for: according to Deviation corrections are applied to the first and second moments respectively; in, This represents the first moment after deviation correction. This represents the second moment after deviation correction; When updating parameters based on the bias-corrected first and second moments, the model training module 74 is used for: according to Conduct the first Step parameter update; in, This indicates the updated parameters. This indicates the parameters before the update. Indicates the learning rate. This represents the regularization term in the denominator to prevent the denominator from being zero.

[0097] In one possible implementation, when processing module 72 performs target detection on the infrared image, locates the transformer region, and crops it to obtain the target infrared image, it is used for: Thermal features of infrared images are extracted using depthwise separable convolution and channel shuffling techniques. According to the thermal characteristics, the multi-scale thermal characteristic fusion is performed through the path aggregation network to generate a multi-resolution feature map; The target center point and the boundary box size in the multi-resolution feature map are predicted to locate the transformer region. The transformer region on the infrared image is cropped to obtain a target infrared image.

[0098] In a possible implementation, after the processing module 72 crops the transformer region on the infrared image to obtain the target infrared image, the processing module is further configured to: The target infrared image is enhanced to obtain an enhanced target infrared image.

[0099] The above transformer fault diagnosis device, through the acquisition module, acquires the infrared image of the transformer, and then the processing module performs target detection on the infrared image, locates the transformer region and performs cropping to obtain a target infrared image, and then the fault recognition module adopts an improved convolutional neural network to perform fault recognition on the target infrared image to obtain a fault category; wherein the L2 pooling layer is used to replace the pooling layer in the original convolutional neural network in the improved convolutional neural network, the L2 pooling layer performs square sum and square root operation on the feature values in the pooling window, avoids the feature detail loss or noise sensitivity problem caused by the traditional pooling method, significantly improves the extraction ability of the transformer infrared image key fault feature, ensures the stability and noise resistance of the feature expression, and adapts to the fault diagnosis demand in complex environment.

[0100] In the embodiment of the application, the Mish activation function is used to combine the characteristics of linearity and nonlinearity, alleviate the gradient vanishing and "dead neuron" problems of the traditional activation function, enhance the modeling ability of the improved convolutional neural network model for the complex nonlinear structure of the infrared image, and improve the convergence speed and the accuracy of fault feature capture. The L2MCNN model fuses L2 pooling, Mish activation, Dropout and GAP, and optimizes the training process through the Adam optimizer and the dynamic learning rate adjustment strategy, and the result shows excellent generalization ability and actual application value.

[0101] In the embodiment of the application, when the target infrared image is acquired, the NanoDet model is used, and the ShuffleNetV2 and PAN network are combined to adapt to the thermal imaging characteristics of the infrared image, quickly and accurately locate the transformer region, significantly reduce the computational complexity and time cost of subsequent processing, and provide technical support for real-time monitoring. Further, the target infrared image is preprocessed through adaptive histogram equalization, bilateral filtering and random rotation, the image quality and data set diversity are effectively improved, and the adaptability of the model to complex scenes is further enhanced, laying a foundation for high-precision fault diagnosis.

[0102] Figure 9 is a schematic diagram of a terminal provided by an embodiment of the application. As shown inFigure 9 As shown, the terminal 9 of this embodiment includes a processor 90, a memory 91, and a computer program 92 stored in the memory 91 and executable on the processor 90. The processor 90 implements the steps in the above method embodiments for diagnosing faults of transformers when executing the computer program 92, for example Figure 1 As shown in steps 101-103. Alternatively, the processor 90 implements the functions of the modules / units in the above apparatus embodiments when executing the computer program 92, for example Figure 7 or the functions of the modules / units shown in FIG. 8.

[0103] For example, the computer program 92 can be segmented into one or more modules / units stored in the memory 91 and executed by the processor 90 to complete the present application. The one or more modules / units can be a series of computer program instruction segments capable of completing a specific function, which are used to describe the execution process of the computer program 92 in the terminal 9. For example, the computer program 92 can be segmented into Figure 7 or the modules / units shown in FIG. 8.

[0104] The terminal 9 can include, but is not limited to, the processor 90 and the memory 91. Those skilled in the art can understand that Figure 9 The terminal 9 shown is only an example and does not limit the terminal 9, which can include more or fewer components than shown, or combine some components, or different components, for example, the terminal can also include an input / output device, a network access device, a bus, etc.

[0105] The processor 90 can be a central processing unit (CPU), and can also be other general-purpose processors, digital signal processors (DSP), application specific integrated circuits (ASIC), field programmable gate arrays (FPGA) or other programmable logic devices, discrete gates or transistor logic components, discrete hardware components, etc. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor.

[0106] The memory 91 can be an internal storage unit of the terminal 9, such as a hard disk or a memory of the terminal 9. The memory 91 can also be an external storage device of the terminal 9, such as a plug-in hard disk, a Smart Media Card (SMC), a Secure Digital (SD) card, a Flash Card, and the like equipped on the terminal 9. Further, the memory 91 can also include both the internal storage unit and the external storage device of the terminal 9. The memory 91 is used to store the computer program and other programs and data required by the terminal. The memory 91 can also be used to temporarily store data that has been output or is to be output.

[0107] Those skilled in the art can clearly understand that, for the convenience and brevity of description, only the division of the above functional units and modules is exemplified, and in actual application, the above functions can be completed by different functional units and modules according to needs, that is, the internal structure of the apparatus is divided into different functional units or modules to complete all or part of the functions described above. Each functional unit and module in the embodiment can be integrated in one processing unit, or each unit can exist physically, or two or more units can be integrated in one unit, and the integrated unit can be realized in the form of hardware or in the form of software functional unit. In addition, the specific names of each functional unit and module are only for the convenience of mutual distinction, and do not limit the protection scope of the present application. The specific working process of the units and modules in the above system can refer to the corresponding process in the foregoing method embodiments, which will not be described here.

[0108] In the above embodiments, the description of each embodiment has its own emphasis, and the parts not described or recorded in detail in a certain embodiment can be referred to the relevant description of other embodiments.

[0109] Those of ordinary skill in the art can appreciate that the units and algorithm steps of each example described in combination with the embodiments disclosed herein can be realized in electronic hardware or in combination of computer software and electronic hardware. Whether the functions are executed in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0110] In the embodiments of the present application, it should be understood that the disclosed apparatus / terminal and method can be implemented in other manners. For example, the embodiments of the apparatus / terminal described above are merely schematic, and the division of the modules or units is merely logical function division, and there can be another division manner in actual implementation. For example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the displayed or discussed mutual couplings or direct couplings or communication connections can be indirect couplings or communication connections through some interfaces, devices or units, and can be electrical, mechanical or in other forms.

[0111] The units described as separate components can or can not be physically separate, and the components shown as units can or can not be physical units, i.e., can be located in one place, or can be distributed on a plurality of network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the embodiments.

[0112] In addition, each functional unit in the various embodiments of the present application can be integrated into a processing unit, or each unit can exist physically, or two or more units can be integrated into one unit. The integrated unit can be implemented in the form of hardware or in the form of a software functional unit.

[0113] The integrated module / unit, if implemented in the form of a software functional unit and sold or used as an independent product, can be stored in a computer readable storage medium. Based on this understanding, all or part of the flow of the above-mentioned embodiment methods can be implemented by a computer program instructing related hardware to complete, and the computer program can be stored in a computer readable storage medium. When the processor executes the computer program, the steps of the above-mentioned transformer fault diagnosis method embodiments can be implemented. The computer program includes computer program code, which can be in the form of source code, object code, executable file or some intermediate form. The computer readable medium can include any entity or device capable of carrying the computer program code, recording medium, U disk, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signal, telecommunication signal and software distribution medium, etc.

[0114] The above-described embodiments are only used to illustrate the technical solutions of the present application, and are not intended to limit the present application; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be included in the protection scope of the present application.

Claims

1. A method of diagnosing a fault of a transformer, characterized by, The method comprises the following steps: acquiring an infrared image of a transformer; performing target detection on the infrared image, locating a transformer region, and performing cropping to obtain a target infrared image; performing fault identification on the target infrared image by using an improved convolutional neural network to obtain a fault category; wherein an L2 pooling layer is used to replace a pooling layer in the original convolutional neural network in the improved convolutional neural network, and the L2 pooling layer performs square sum and square root operations on feature values in a pooling window.

2. The transformer fault diagnostic method according to claim 1, characterized by, The method of performing fault identification on the target infrared image by using the improved convolutional neural network to obtain the fault category comprises the following steps: inputting the target infrared image into the improved convolutional neural network, performing convolution operation in the multiple convolution layers, extracting multi-dimensional features of the target infrared image, and outputting a first feature map with the same size as the target infrared image; performing non-linear transformation on the features in the first feature map by using a Mish function in a Mish activation layer to output a second feature map, wherein the second feature map has the same size as the first feature map; performing L2 pooling processing on local input vectors of the second feature map in an L2 pooling layer to output a third feature map, wherein the third feature map is a compressed feature map of the second feature map; randomly discarding a preset number of neurons in a Dropout layer to output a fourth feature map, wherein the fourth feature map has the same size as the third feature map; performing global pooling on the dimensions of the current input feature map in a global average pooling layer, and reducing each channel spatial dimension to 1 to output a feature vector with a spatial dimension of 1*1; mapping the feature vector with the spatial dimension of 1*1 to a preset dimension in a fully connected layer to realize the mapping of the features to the fault category, and outputting a feature vector with the preset dimension; converting the feature vector with the preset dimension into a probability distribution in a Softmax layer to output the prediction probability of each category; determining the final fault category according to the prediction probability of each category in a classification layer and outputting the final fault category.

3. The transformer fault diagnostic method according to claim 2, characterized by, The method of performing L2 pooling processing on the local input vectors of the second feature map comprises the following steps: According to perform L2 pooling processing on the local input vector of the second feature map; wherein, denotes the feature vector after L2 pooling processing, denotes the local input vector, , denotes a positive integer, and is the number of local input vectors.

4. The transformer fault diagnostic method according to claim 2, characterized by, The Mish function is as follows: ； wherein, denotes an output result of a Mish function, denotes a feature in the first feature map, denotes a hyperbolic tangent function, denotes a natural constant.

5. The transformer fault diagnostic method according to claim 2, characterized by, Before the method of performing fault identification on the target infrared image by using the improved convolutional neural network to obtain the fault category, the method further comprises the following steps: training the improved convolutional neural network to obtain a trained improved convolutional neural network; in the process of training the improved convolutional neural network, the first moment and the second moment are used to adaptively adjust the learning rate for each parameter; the first moment and the second moment are respectively subjected to bias correction; and parameter updating is performed according to the bias-corrected first moment and the bias-corrected second moment.

6. The transformer fault diagnostic method according to claim 5, characterized by, The method of using the first moment and the second moment to adaptively adjust the learning rate for each parameter comprises the following steps: According to Adapt the learning rate for each parameter; in, Indicates the first The first moment of the step, Indicates the first The second moment of the step, This represents the exponential decay rate of the first moment. This represents the exponential decay rate of the second moment. Indicates the first The first moment of the step, Indicates the first The second moment of the step, Indicates the first The gradient of the step; The method of respectively subjecting the first moment and the second moment to bias correction comprises the following steps: According to bias correction is performed on the first moment and the second moment, respectively; wherein, denotes the first moment after bias correction, denotes the second moment after bias correction; The method of performing parameter updating according to the bias-corrected first moment and the bias-corrected second moment comprises the following steps: According to performing the parameter update of step ; wherein, denotes the updated parameter, denotes the updated parameter, denotes the learning rate, denotes a regularization term for the denominator to prevent the denominator from being 0.

7. The transformer fault diagnostic method according to any one of claims 1 to 6, characterized by, The method of performing target detection on the infrared image, locating the transformer region, and performing cropping to obtain the target infrared image comprises the following steps: extracting thermal features of the infrared image by using a depth separable convolution technology and a channel shuffle technology; According to the thermal features, multi-scale thermal feature fusion is performed through a path aggregation network to generate a multi-resolution feature map; A target center point and a bounding box size in the multi-resolution feature map are predicted to locate a transformer region; The transformer region on the infrared image is cropped to obtain a target infrared image.

8. The transformer fault diagnostic method according to claim 7, characterized by, After the transformer region on the infrared image is cropped to obtain the target infrared image, the method further includes: The target infrared image is subjected to enhancement processing to obtain an enhanced target infrared image.

9. A fault diagnosis device for a transformer, characterized in that, The method includes: An acquisition module is configured to acquire an infrared image of a transformer; A processing module is configured to perform target detection on the infrared image, locate a transformer region, and crop the transformer region to obtain a target infrared image; A fault identification module is configured to perform fault identification on the target infrared image by using an improved convolutional neural network to obtain a fault category; wherein an L2 pooling layer is used to replace a pooling layer in the original convolutional neural network in the improved convolutional neural network, and the L2 pooling layer performs square sum and square root operations on feature values in a pooling window.

10. A terminal comprising a memory for storing a computer program and a processor for invoking and running the computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method according to any one of claims 1 to 8.

Images