Power equipment partial discharge infrared image generation method

Abstract

The invention discloses a power equipment partial discharge infrared image generation method, and aims to solve the problems that samples are insufficient and generated image detail features are lost when power equipment has a partial discharge fault. The input visible light image and the partial discharge infrared image of the power equipment are combined, the combined image is input into the generator to obtain a fused image, the fused image is input into the discriminator, and the discriminator distinguishes and identifies the fused image and the visible light image of the power equipment. Wherein the generator is composed of an encoder and a decoder, the encoder is composed of five MDCN modules, the five MDCN modules are connected through a CBAM-DenseNet network, and the network structure of the decoder is composed of three convolutional layers; the method for generating the partial discharge infrared image of the power equipment can accurately enhance and select features such as edges and textures in the visible light image, and improves the expression capability of the model for beneficial features.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of partial discharge defect detection of power equipment, and particularly relates to a method for generating an infrared image of partial discharge of power equipment. BACKGROUND

[0002] Stable operation of power equipment is an important basis for ensuring safe operation of the power system. When the power equipment is in operation, discharge may occur inside or on the surface of the insulating material, which generates local high temperature in the equipment, leading to aging of the insulating material. Monitoring the heat change is very important for preventing equipment failure. Infrared image detection can be used to detect the heat generated by partial discharge of power equipment. By using an infrared thermal imager, the temperature distribution on the surface of the power equipment and its surroundings can be captured, so that possible abnormal heat sources or hot spot areas can be identified, which helps to discover potential equipment problems in time and to maintain and repair in time. However, due to the small number of infrared fault samples and the fact that most data sets are not public, it is difficult to train a detection model with good performance. Therefore, the method for generating and fusing visible light and infrared images based on the generative adversarial network can not only expand the number of samples, but also enhance the edge and texture information in the samples, so as to expand the data set.

[0003] The GAN (Generative Adversarial Network) is composed of a generator G and a discriminator D. The generator is used to continuously generate samples, and the discriminator is used to distinguish whether the samples generated by the generator are generated images or real images. The two continuously compete until the discriminator cannot distinguish between true and false. Since the heat radiation in the visible light image and the infrared image is two different phenomena, the FusionGAN (FusionGAN) is the first to propose a generative adversarial network that can simultaneously maintain the texture information of the visible light image and the heat radiation information of the infrared image on the basis of the GAN, and solve the image fusion task.

[0004] However, although the existing generative infrared and visible light image fusion adversarial network model has achieved good fusion effect, in the process of image feature extraction, only a single scale convolution layer is used to extract features from the image, and in the process of image fusion, the edge features and multi-scale information of the source image are lost, resulting in loss of detailed features of the generated image. SUMMARY

[0005] The present application provides a method for generating an infrared image of partial discharge of power equipment, which aims to solve the problem of loss of detailed features in generating a fused image in the prior art.

[0006] A method for generating an infrared image of partial discharge of power equipment, comprising the following steps:

[0007] Step 1, collecting an initial image data set of the power equipment, the initial image including a visible light image and a partial discharge infrared image of the power equipment, preprocessing the visible light image and the partial discharge infrared image of the power equipment in the initial image data set, and dividing into a training set and a test set;

[0008] Step 2, inputting the training set into a generative adversarial network for training, the generative adversarial network including a generator and a discriminator, combining the input visible light image and the partial discharge infrared image of the power equipment, inputting the combined image into the generator to obtain a fusion image, inputting the fusion image into the discriminator, and the discriminator distinguishing and identifying the fusion image from the visible light image of the power equipment, the generator being composed of an encoder and a decoder, the combined visible light image and the partial discharge infrared image of the power equipment being input into the encoder, the encoder extracting features of the combined image to obtain a fusion feature image, and the fusion feature image being input into the decoder to obtain the fusion image, wherein the generator is composed of the encoder and the decoder, the encoder is composed of five MDCN modules and connected through a CBAM-DenseNet network, and the decoder network structure is composed of three convolutional layers;

[0009] Step 3, inputting the test set into the trained generator to obtain a final fusion image. BRIEF DESCRIPTION OF DRAWINGS

[0010] Figure 1 is a flowchart of the power equipment partial discharge infrared image generation method provided by the embodiment of the present application;

[0011] Figure 2 is a structure diagram of the generator of the generative adversarial network in the present application;

[0012] Figure 3 is Figure 1 is a schematic diagram of the MDCN module in the generator of the generative adversarial network shown in the figure;

[0013] Figure 4 is a structure diagram of the discriminator of the generative adversarial network in the present application. DETAILED DESCRIPTION

[0014] In order to make the purpose, technical scheme and advantages of the embodiments of the present application more clear, the technical scheme in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0015] Please refer to Figure 1, an embodiment of the present invention provides a device image generation method, the method comprising steps 1 to 3:

[0016] Step 1: Acquire an initial image dataset of the power equipment, wherein the initial images include visible light images and partial discharge infrared images of the power equipment, preprocess the visible light images and partial discharge infrared images of the power equipment in the initial image dataset, and divide them into a training set and a test set;

[0017] Step 2: Input the training set into a generative adversarial network for training. The generative adversarial network includes a generator and a discriminator. The input visible light image of the power equipment and the partial discharge infrared image are combined. The combined image is input into the generator to obtain a fused image. The fused image is input into the discriminator, and the discriminator distinguishes and identifies the fused image from the visible light image of the power equipment.

[0018] like Figure 2 As shown, step 2 specifically includes:

[0019] The generator consists of an encoder and a decoder. The input visible light image of the power equipment and the partial discharge infrared image are combined and input into the encoder. The encoder extracts features from the combined image to obtain a fused feature image, which is then input into the decoder to obtain a fused image.

[0020] The encoder consists of five MDCN modules connected through a CBAM-DenseNet network.

[0021] In order to accurately enhance and select features such as edges and texture features in visible light images and improve the model's ability to express beneficial features, the CBAM-DenseNet network adds the CBAM module before and after the ReLU activation function.

[0022] like Figure 3 As shown in the figure, the MDCN module in the encoder first uses a convolution with a 1×1 convolution kernel to compress the number of channels and reduce the network parameters.

[0023] Feature extraction is performed after compressing the number of channels. Multi-scale convolution kernels are used to enhance the feature extraction capability. Each branch of the module uses deformable convolution with kernel sizes of 3×3, 5×5, and 7×7 to extract feature information of different scales, and then the feature information of different scales is fused and spliced.

[0024] Then, a convolution with a 1×1 kernel is used to expand the number of channels.

[0025] Finally, the compressed and expanded features are added together to obtain the fused features.

[0026] Specifically, the working process of the MDCN module is as follows:

[0027]

[0028] wherein, F1, F2 and F3 represent different scale feature maps obtained, DConv 3×3 (), DConv 5×5 () and DConv 7×7 () represent deformable convolution with convolution kernel of 3x3, 5x5 and 7x7, Conv 1×1-1 () and Conv 1×1-2 () represent convolution with convolution kernel of 1x1, D represents the splicing operation on different scale feature maps, W1, W2, W3, W4 and W5 represent the weights in each operation, represents the input of the nth MDCN module, represents the output of the nth MDCN module.

[0029] The decoder network structure is composed of three convolutional layers, the first two convolutional layers are composed of convolution with convolution kernel of 3x3, batch normalization and LRelu activation function, and the third layer is composed of convolution with convolution kernel of 3x3, batch normalization and Tanh activation function.

[0030] As shown in Figure 4 , in step 2, the discriminator is composed of five convolutional layers, the first four layers use convolution with convolution kernel of 3x3 and LRelu activation function, and batch normalization is added in the second to fourth layers, and the last layer uses a fully connected layer and tanh activation function to distinguish whether the input image is a fusion image or a visible light image.

[0031] The generator loss function L G is:

[0032] L G = E(ln(1-D(I f )))+λL content

[0033]

[0034] wherein, E(ln(1-D(I f ))) represents the adversarial loss of the generator and the discriminator, L content represents the content loss, which is composed of three parts of intensity loss, gradient loss and similarity loss, λ represents the weight parameter for balancing the adversarial loss and the content loss, I f represents the generated fusion image, I r represents the partial discharge infrared image, I v represents the visible light image, D(I f) represents the probability of generating the fusion image, H and W represent the height and width of the input picture, represents the intensity loss, constrains the infrared thermal radiation information caused by partial discharge, represents the gradient loss, preserves the edge and texture information of the power equipment visible light image, the similarity loss constrains the contrast, brightness and structure of the input image and the fusion image, L SSIM (I f ,I v ) represents the similarity between the fusion image and the power equipment visible light image, L SSIM (I f ,I r ) represents the similarity between the fusion image and the partial discharge infrared image, and α, β and γ represent the weight parameters for controlling the three loss parts.

[0035] The discriminator loss function L D is:

[0036] L D = E[-ln(D(I v ))] + E[-ln(1-D(I f ))]

[0037] Wherein, E[-ln(D(I v ))] represents the loss of the discriminator inputting the power equipment visible light image, and E[-ln(1-D(I f ))] represents the loss of the discriminator inputting the fusion image.

[0038] In step 3, the test set is input into the trained generator to obtain the final fusion image.

Claims

1. A method for generating infrared images of partial discharge of power equipment, characterized in that: The following steps are involved: Step 1: Acquire an initial image dataset of the power equipment, wherein the initial images include visible light images and partial discharge infrared images of the power equipment, preprocess the visible light images and partial discharge infrared images of the power equipment in the initial image dataset, and divide them into a training set and a test set; Step 2: Input the training set into a generative adversarial network for training. The generative adversarial network includes a generator and a discriminator. The input visible light image of the power equipment and the partial discharge infrared image are combined, and the combined image is input into the generator to obtain a fused image. The fused image is input into the discriminator, and the discriminator distinguishes and identifies the fused image from the visible light image of the power equipment. Step 3: Input the test set into the trained generator to obtain the final fused image.

2. The method for generating infrared images of partial discharge of power equipment according to claim 1, characterized in that: In step 2, the generator consists of an encoder and a decoder. The input visible light image of the power equipment and the partial discharge infrared image are combined and input into the encoder. The encoder extracts features from the combined image to obtain a fused feature image. The fused feature image is input into the decoder to obtain a fused image.

3. The method for generating infrared images of partial discharge of electric power equipment according to claim 2, characterized in that: The encoder consists of five MDCN modules connected by a CBAM-DenseNet network; The CBAM-DenseNet network adds the CBAM module before and after the ReLU activation function, which can accurately enhance and select features such as edge and texture features in visible light images, improving the model's ability to express useful features.

4. The method for generating infrared images of partial discharge of electric power equipment according to claim 3, characterized in that: The MDCN module uses a convolution with a convolution kernel of 1×1 to compress and expand the number of channels. First, the number of channels is compressed to reduce network parameters. After feature extraction, the number of channels is expanded. Feature extraction uses a multi-scale convolution kernel to enhance feature extraction capabilities. Each branch of the module uses deformable convolution with convolution kernel sizes of 3×3, 5×5, and 7×7 to extract feature information of different scales. The feature information of different scales is then fused and spliced. Finally, the compressed and expanded features are added to obtain the fused features. The process is as follows: D=[DConv 3×3 ,DConv 5×5 ,DConv 7×7 ] Among them, F1, F2 and F3 represent the feature maps of different scales obtained, DConv 3×3 (), DConv 5×5 () and DConv 7×7 () indicates deformable convolution with convolution kernels of 3×3, 5×5 and 7×7, Conv 1×1-1 () and Conv 1×1-2 () represents a convolution with a convolution kernel of 1×1, D represents the concatenation operation of feature maps of different scales, W1, W2, W3, W4 and W5 represent the weights in each step of the operation, represents the input of the nth MDCN module, Represents the output of the nth MDCN module.

5. The method for generating infrared images of partial discharge of electric power equipment according to claim 4, characterized in that: The decoder network structure consists of three convolutional layers, the first two convolutional layers consist of convolution with a convolution kernel of 3×3, batch normalization and LRelu activation function, and the third layer consists of convolution with a convolution kernel of 3×3, batch normalization and Tanh activation function.

6. The method for generating infrared images of partial discharge of electric power equipment according to claim 5, characterized in that: In step 2, the discriminator consists of five convolutional layers. The first four layers use convolution with a convolution kernel of 3×3 and LRelu activation function, and batch normalization is added to the second to fourth layers. The last layer uses a fully connected layer and tanh activation function to discriminate whether the input image is a fused image or a visible light image.

7. The method for generating infrared images of partial discharge of electric power equipment according to claim 6, characterized in that: The generator loss function L G for: L G =E(ln(1-D(I f )))+λL content Among them, E(ln(1-D(I f ))) represents the adversarial loss between the generator and the discriminator, L content Represents content loss, which consists of three parts: intensity loss, gradient loss, and similarity loss. λ represents the weight parameter used to balance adversarial loss and content loss. f represents the generated fusion image, I r Represents the infrared image of partial discharge, I v Represents the visible light image, D(I f ) represents the probability of generating a fused image, H and W represent the height and width of the input image, Indicates the intensity loss and constrains the infrared thermal radiation information caused by partial discharge. Represents the gradient loss, which retains the edge and texture information of the visible light image of the power equipment. The similarity loss constrains the contrast, brightness and structure of the input image and the fused image. SSIM (I f ,I v ) represents the similarity between the fused image and the visible light image of the power equipment, L SSIM (I f ,I r ) represents the similarity between the fused image and the PD infrared image, α, β and γ represent the weight parameters controlling the three-part loss; Discriminator loss function L D for: L D =E[-ln(D(I v ))]+E[-ln(1-D(I f ))] Among them, E[-ln(D(I v ))] represents the visible light image loss of the discriminator input power device, E[-ln(1-D(I f ))] represents the discriminator input fusion image loss.

Images