Statistical method for determining the number of samples for a transformer key component based on the knowledge of the regulation

Abstract

The application discloses a transformer key component effective sample expansion method based on reduction knowledge, and comprises the following steps: A, constructing an image data set of a transformer key component; B, inputting the image data set into a semantic enhancement network ISDA to perform image data expansion; C, creating an image compensation network based on a guided back propagation, and compensating image data expanded in the step B; and D, mixing the image data set compensated in the step C with a real inspection image data set to obtain a mixed image data set, and completing effective sample expansion. The application can improve the defects of the prior art, effectively increase the richness and diversity of transformer component image samples, and meet the requirements of a detection algorithm from a sample level.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of transformer detection, and in particular to a method for expanding effective samples of key components of a transformer based on protocol knowledge. BACKGROUND

[0002] Currently, there is no public data set for detecting key components of a transformer, and in order to deeply study and solve the problem of detecting oil leakage of a transformer, it is necessary to first complete the detection of different key components of a transformer. Due to large changes in the visual angle and visual distance of inspection, the sample collection of different components of a transformer is not comprehensive, and there is a sample imbalance problem. Traditional image expansion focuses on expansion in quantity, does not fully exploit the characteristics of the sample itself, and does not utilize existing prior knowledge, which restricts the development of the method for expanding effective samples of key components of a transformer. Specifically, the traditional method for expanding image samples is to perform geometric transformation on the original image, such as obtaining new samples by translation, flipping, and elastic deformation, etc. to alleviate the problems of insufficient samples and overfitting, and then deriving new sample acquisition methods such as rectangular erasing and image fusion. Among them, traditional data expansion methods such as cropping, flipping, and rotating do not involve semantic information transformation, and cannot play a real role in data expansion. Generative Adversarial Networks (GAN) is a data expansion method based on semantic information, and based on this, a series of GAN-based image data expansion models have been derived. Image deep feature interpolation is also an image processing method, which needs to manually find the direction of deep feature interpolation in the feature space, and the workload of this process is huge, and only a very limited number of directions can be found by manual human means. SUMMARY

[0003] The technical problem to be solved by the present application is to provide a method for expanding effective samples of key components of a transformer based on protocol knowledge, which can overcome the shortcomings of the prior art and effectively increase the richness and diversity of image samples of components of a transformer, and meet the needs of detection algorithms at the sample level.

[0004] To solve the above technical problems, the technical solutions adopted by the present application are as follows.

[0005] A method for expanding effective samples of key components of a transformer based on protocol knowledge, comprising the following steps:

[0006] A. Constructing an image data set of key components of a transformer;

[0007] B. Inputting the image data set into a semantic enhancement network ISDA to expand the image data;

[0008] C. Creating an image compensation network based on guided backpropagation to compensate for the image data expanded in step B.

[0009] D, the compensated image data set after step C is mixed with the real inspection image data set to obtain a mixed image data set, and effective sample expansion is completed.

[0010] As preferred, in step A, the transformer key components are divided into six categories: transformer body, oil conservator, bushing, tap changer, radiator and relay.

[0011] As preferred, in step B, the semantic enhancement network ISDA adopts a residual network, and an image edge sharpening module is added in the semantic enhancement network ISDA, the image edge sharpening module detects edge information of the image by calculating the gradient, and sharpens the edge pixels, and the calculation formula is,

[0012]

[0013] Wherein, f(x, y) and l(x, y) are two images, C is a constant to control the image sharpening degree, and e(x, y) is an image containing final edge information, for each pixel on the image, if e(x, y)>0, it indicates that the pixel is an edge and needs to be sharpened, if e(x, y)=0, it indicates that the pixel is not an edge and does not need to be sharpened, and the pixel value remains unchanged.

[0014] As preferred, in step B, the covariance matrix of the data is estimated first, so as to capture the change direction of the variance of each type of data, and then the sampling direction is obtained from the normal distribution, so as to realize the expansion of the data set.

[0015] As preferred, in step C, a deep separable dilated convolution pyramid guided back propagation network is constructed, the response intensity matrix R of the artificial image is calculated, the Hadamard product of the pixel matrix of the artificial scene image is calculated, the normalized processing is performed to obtain the compensated transformer key component image, the deep separable dilated convolution pyramid is composed of deep dilated convolution and point convolution, and finally the convolution results are summed to obtain the final feature map.

[0016] The beneficial effects brought by the above technical scheme are that: different from the traditional expansion method which focuses on the expansion of the number of samples, the present application proposes a transformer key component effective sample expansion method based on reduction knowledge, realizes the generation of regular artificial images by using reduction knowledge, and performs optimization compensation, effectively increases the richness and diversity of the power component image samples, meets the needs of the detection algorithm from the sample level, has great scientific value, and further provides rich data resources for researchers in the same field, promotes the improvement of the algorithm model performance and the popularization of the practical application. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1is the statistical distribution of the number of 6 kinds of wind key component identification data samples collected by the inventor.

[0018] Figure 2 is the overall flowchart of the present application.

[0019] Figure 3 is the flowchart of the image data augmentation of the semantic enhancement network ISDA in the present application.

[0020] Figure 4 is the flowchart of the image compensation of the image compensation network based on guided back propagation in the present application. DETAILED DESCRIPTION

[0021] Referring to Figures 1-4 , one embodiment of the present application includes the following steps,

[0022] A. According to the monograph "Application Technology of Large Oil-immersed Power Transformer" (Dong Baohua, China Electric Power Press, 2014) and DL / T573-2021 standard, the 6 kinds of key components related to transformer oil leakage detection are determined as transformer body, oil pillow, bushing, tap changer, radiator and relay; the three-dimensional modeling software is used to standardize modeling of the 6 kinds of key components, to generate artificial scene images of different angles and shapes; in the modeling process, the virtual visual camera is configured, the light source, depth of field and other settings are used to increase the reality effect of the three-dimensional model, and the diversity of the transformer key components is increased by changing the above settings; through scaling, rotating and other operations, the virtual camera is used to shoot to obtain diversity artificial key component images of different angles and different proportions, so as to construct the image dataset of the transformer key components; through Figure 1 It can be seen that there is a problem of unbalanced sample distribution;

[0023] B. The semantic enhancement network adopts a residual network, in order to solve the problem of low resolution of the generated image, an image edge sharpening module is added in the semantic enhancement network, the image edge sharpening module detects the edge information of the image by calculating the gradient, and sharpens the edge pixels, the main purpose is to compensate the image contour and highlight the edge information of the image to make the image clearer; the calculation formula is,

[0024]

[0025] Wherein, f(x, y) and l(x, y) are two images, C is a constant to control the image sharpening degree, and e(x, y) is an image containing final edge information, for each pixel on the image, if e(x, y) > 0, it indicates that the pixel is an edge and needs to be sharpened, if e(x, y) = 0, it indicates that the pixel is not an edge and does not need to be sharpened, the pixel value remains unchanged;

[0026] The semantic enhancement network can find data features belonging to the same category but different semantics, and clearly shows the direction of sampling, avoiding the direction of sampling without practical significance. In the sampling process, instead of randomly and evenly sampling in all directions, the covariance matrix of the data is first estimated to capture the change direction of the variance of each category of data, and then the sampling direction is sampled from the normal distribution to achieve the expansion of the data set.

[0027] C, a deep separable dilated convolution pyramid guided backpropagation network is constructed, the response intensity matrix R of the artificial image is calculated, the Hadamard product of the pixel matrix of the artificial scene image is calculated, and the compensated transformer key component image is obtained after normalization processing, in order to reduce the calculation overhead of the control guided backpropagation compensation network, a deep separable dilated convolution pyramid structure is constructed, which is composed of a deep dilated convolution and a point convolution. The calculation method of deep dilated convolution is very simple, and deep dilated convolution can expand the receptive field without changing the convolution kernel. It performs dilated convolution on each channel of the input feature map, where rate represents the dilated rate, and finally the convolution result is summed to obtain the final feature map; the point convolution is actually a 1*1 convolution, which mainly has two functions. The first function is to adjust the expected output channel number, and the second function is to fuse the feature maps output by the deep dilated convolution;

[0028] D, the image data set compensated by step C is mixed with the real inspection image data set to obtain a mixed image data set, and the effective sample expansion is completed.

[0029] In the description of the present application, it should be understood that the orientations or positional relationships indicated by the terms "longitudinal", "transverse", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer" and the like are based on the orientations or positional relationships shown in the drawings, and are only for the convenience of describing the present application, and do not indicate or imply that the devices or elements referred to must have a particular orientation, be constructed and operated in a particular orientation, therefore cannot be understood as a limitation of the present application.

[0030] The basic principles and main features of the present application and the advantages of the present application are shown and described above. Those skilled in the art should understand that the present application is not limited by the above examples, the above examples and descriptions in the specification are only to illustrate the principles of the present application, and various changes and improvements can be made to the present application without departing from the spirit and scope of the present application, and these changes and improvements all fall within the scope of the claimed present application. The scope of protection of the present application is defined by the appended claims and their equivalents.

Claims

1. A method for expanding the effective sample of key transformer components based on specification knowledge, characterized in that... Includes the following steps: A. Construct an image dataset of key transformer components; B. Input the image dataset into the semantic augmentation network ISDA for image data augmentation, specifically: The Semantic Augmentation Network (ISDA) employs a residual network. An image edge sharpening module is added to ISDA. This module detects edge information by calculating gradients and sharpens edge pixels. The calculation formula is as follows: Where f(x,y) and l(x,y) are two images, C is a constant used to control the degree of image sharpening, and e(x,y) is an image containing the final edge information. For each pixel in the image, if e(x,y)>0, it means that the pixel is an edge and needs to be sharpened. If e(x,y)=0, it means that the pixel is not an edge and does not need to be sharpened. The pixel value remains unchanged. First, the covariance matrix of the data is estimated to capture the direction of variance change for each class of data. Then, the direction is sampled from the normal distribution to expand the dataset. C. Create an image compensation network based on guided backpropagation to compensate for the image data amplified in step B, specifically as follows: A depthwise separable dilated convolution pyramid guided backpropagation network is constructed. The response intensity matrix R of the artificial image is calculated and the Hadamard product of the pixel matrix of the artificial scene image is obtained. After normalization, the compensated image of the key components of the transformer is obtained. The depthwise separable dilated convolution pyramid is composed of two parts: depthwise dilated convolution and point convolution. Finally, the convolution results are merged to obtain the final feature map. D. Mix the image dataset after compensation in step C with the real inspection image dataset to obtain a mixed image dataset, thus completing the effective sample expansion.

2. The method for expanding the effective sample of key transformer components based on specification knowledge according to claim 1, characterized in that: In step A, the key components of the transformer are divided into six categories: transformer body, oil conservator, bushing, tap changer, radiator and relay.

Images