GIS equipment SF6 leakage identification method based on component and SF6 infrared feature image fusion

Abstract

A GIS equipment SF6 leakage identification method based on component and SF6 infrared feature image fusion combines infrared image processing with a computer vision principle, carries out grey-scale mapping by using an obtained visible light image and an obtained infrared image, and then decomposes a source image through a Gaussian filtering image decomposition method, the method is sensitive to illumination change and is used for acquiring features of GIS equipment components. And the infrared image can reveal thermal information insensitive to illumination change but cannot provide enough texture details, and is used for identifying SF6 infrared features. The infrared image and the visible light image are fused through a multi-layer fusion strategy, so that the SF6 gas leakage image quality of the GIS equipment is improved, and the distinguishing degree of the image is enhanced. Through qualitative evaluation and quantitative evaluation of the image, it can be proved that the method effectively improves the intelligibility of the SF6 gas leakage detection image of the GIS equipment, and the background interference problem of equipment image features can be solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of SF6 leakage identification of electric power GIS equipment, and in particular relates to a GIS equipment SF6 leakage identification method based on component and SF6 infrared feature image fusion. Background Art

[0002] GIS equipment has been widely used in power systems in recent years due to its advantages such as small footprint and stable operation. However, SF6 gas leakage may occur during GIS operation. On the one hand, SF6 leakage will reduce the insulation strength of the equipment, making the equipment unable to operate stably; on the other hand, leakage will seriously damage the atmospheric environment. Infrared imaging leak detection has the advantages of non-contact, high sensitivity, and mature technology. It can detect SF6 leakage of equipment in a timely and effective manner to prevent equipment failure. However, a large number of inspections or online monitoring still require manual analysis, which is time-consuming and inefficient, and it is difficult to cope with the large number of infrared images generated by the inspection. Therefore, researchers have conducted a series of studies on infrared image recognition. By using different types of features or classification models to analyze and process images, GIS-related equipment is then determined and identified. With the development of artificial intelligence technology, image detection algorithms based on deep learning have been widely used in the field of visible light object detection. However, they are rarely used in infrared image recognition and detection. Therefore, it is urgent to develop GIS-related equipment features based on infrared images and SF6 gas leakage identification methods. Summary of the invention

[0003] The technical problem to be solved by the present invention is to provide a GIS equipment SF6 leakage identification method based on component and SF6 infrared feature image fusion, and to fuse infrared images and visible light images through a multi-layer fusion strategy to improve the quality of SF6 gas leakage images and enhance the image recognizability.

[0004] In order to solve the above technical problems, the technical solution adopted by the present invention is: The SF6 leakage identification method based on the fusion of GIS equipment component features and SF6 infrared feature images is as follows: Step 1: Image separation: Step 1.1, construct grayscale image: the source image of SF6 gas leakage is a color image, which is converted into a grayscale image before image separation; Step 1.2, decompose the source image: decompose the source image into a base layer image and an original image; once the base layer image is obtained, the detail layer image, i.e., the component features and possible SF6 infrared features contained therein, can be obtained by subtracting the basic part from the source image; Step 2: Image fusion: Step 2.1, pre-fusion base layer construction: The base parts extracted from the source image contain common features and redundant information, and a weighted average strategy is selected to fuse these base parts; Step 2.2: Pre-fusion detail layer construction: The detail parts separated from the source image are processed and then newly fused. Preferably, in step 1.1, the color image is stored in a computer in the form of a three-dimensional matrix of M×N×3, where M represents the number of rows of the image and N represents the number of columns of the image, and after processing, it is converted into a two-dimensional matrix of M×N for storage; the conversion method and its calculation formula are as follows: (1); Where α1, α2 and α3 represent the weights of pixels in PR, PG and PB, respectively.

[0005] Preferably, in step 1.2, it is assumed that the visible light image and the infrared image of the obtained source image are P1 and P2 respectively; the source image is decomposed into two scale representations, wherein the base layer image is obtained by solving the following formula: (2); In the formula, is the base layer image, is the original image, is the convolution operation, G is the convolution kernel; is the detail layer image; once the base layer image is obtained, the detail layer image can be obtained by subtracting the basic part from the source image; (3); Source image decomposition aims to separate each source image into a base layer image containing the basic frame and a detail layer image containing small-scale details; The source images are denoted as P1 and P2; the base layer image of each source image is obtained by solving equations (2) and (3) and detail layer images , where n∈{1,2}.

[0006] Preferably, in step 2.1, the base parts extracted from the source image contain common features and redundant information; a weighted average strategy is selected to fuse the contents of these base parts; and the fused base layer is calculated by equation (4): (4); Where (x, y) represents the image intensity at , and The corresponding position in ; and Respectively and The weight value of the pixel in .

[0007] Preferably, in step 2.2, first obtain its depth feature , represents the feature map of the nth detail content extracted by the kth layer, and m is the number of channels of the kth layer; (5); In the formula, represents an activation layer in the VGG network, let , indicating that four activation layers are selected from the VGG network as the source of deep features; , ; set up Represents the position in the feature map Department content, is an M-dimensional vector; The l1 norm of can be the activity level measure of the source detail content, and the initial activity level map is obtained by the following equation: (6); Based on the initial activity level map, we use the block-based average operator to calculate and obtain a robust activity level map; (7); In the formula, q determines the block size; Getting an activity level graph After that, the soft maximum operator is used to calculate the initial weight map: (8); Where N=2, Represents the initial weight mapping value in the range [0,1]; After obtaining each initial weight map Afterwards, the weight map is resized to the size of the input detail content using an upsampling operator: (9); Through the above formula, four pairs of weight mapping sets can be obtained; based on each pair of mapping sets, the following equation is used to calculate the alternative fusion detail layer: (10); Finally, the maximum value is selected from the four initial fused detail contents at each position to obtain the fused detail content. ; Use a fused pre-fused base layer and fused pre-fused detail layers Add them together to reconstruct the final fused image.

[0008] The present invention can achieve the following beneficial effects: The present invention aims at the problems that the GIS equipment of the power system is complex and changeable, and SF6 gas leakage may occur during operation, which requires a large number of photo inspections or online monitoring, and the subsequent analysis and processing of a large number of images. This method combines infrared image processing and computer vision principles, uses the acquired visible light image and infrared image to perform grayscale mapping, and then decomposes the source image through the Gaussian filter image decomposition method: the visible light image contains rich texture details, but is very sensitive to lighting changes, and is used to obtain the characteristics of GIS equipment components; and the infrared image can reveal thermal information that is insensitive to lighting changes but cannot provide sufficient texture details for the identification of SF6 infrared features. The infrared image and the visible light image are fused through a multi-layer fusion strategy to improve the quality of the SF6 gas leakage image and enhance the recognizability of the image. Through qualitative and quantitative evaluation of the image, it can be proved that this method effectively improves the intelligibility of the SF6 gas leakage detection image and helps to solve the background interference problem of the equipment image features. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] The present invention will be further described below in conjunction with the accompanying drawings and embodiments: Figure 1 It is a source image decomposition framework diagram of the present invention; Figure 2 This is a valve leakage image fusion framework diagram of the present invention; Figure 3 It is a framework diagram for integrating the details of the present invention; Figure 4 This is the image comparison result diagram of the present invention; Figure 5 This is a quantitative comparison chart of the performance of the present invention. DETAILED DESCRIPTION

[0010] The preferred solution is Figures 1 to 5 As shown in the figure, the GIS equipment SF6 leakage identification method based on the fusion of components and SF6 infrared feature images includes the following steps: Step 1: Image separation: GIS equipment components with SF6 gas leakage are detected by infrared detectors and online monitoring cameras, and the obtained infrared images and visible light images are preprocessed. The three-dimensional image is converted into a two-dimensional image using grayscale conversion, and then the source image is decomposed into a basic part and a detailed part by a dual-scale image decomposition method. The deep features of the detailed content, i.e. the contained component features and the possible contained SF6 infrared features, are extracted to provide a basis for the next step of image fusion. The specific steps are as follows: Step 1.1, construct grayscale image: The source image of SF6 gas leakage is a color image, which needs to be converted into a grayscale image before image separation. The output format of the color image is RGB, which is saved in the computer in the form of a three-dimensional matrix of M×N×3, where M represents the number of rows of the image and N represents the number of columns of the image. After processing, it is converted into a two-dimensional matrix of M×N for storage. The conversion method and its calculation formula are as follows: (1); Where α1, α2 and α3 represent the weights of pixels in PR, PG and PB respectively, which are summarized based on the experience of human brightness perception system, and α1 = 0.2989, α2 = 0.5870 and α3 = 0.1140 are taken.

[0011] Step 1.2, decompose the source image: Gas leakage is mostly manifested as local brightness or temperature changes in the image. The average filter may be too smooth, resulting in the image losing some details and being unclear. This method uses a shorter and more efficient image decomposition method, which can retain more edge and detail information while removing noise, helping to highlight these changes. Assume that the visible light image and infrared image of the obtained source image are P1 and P2 respectively. Figure 1 As shown, the source image is decomposed into two scale representations, where the base layer image is obtained by solving the following formula: (2); In the formula, is the base layer image, is the original image, is the convolution operation, and G is the convolution kernel. is the detail layer image. Once the base layer image is obtained, the detail layer image can be obtained by subtracting the base part from the source image.

[0012] (3); Source image decomposition aims to separate each source image into a base layer image containing the basic frame and a detail layer image containing small-scale details.

[0013] like Figure 1 As shown, the source images are represented as P1 and P2. First, the base layer image of each source image is obtained by solving equations (2) and (3): and detail layer images , where n∈{1,2}.

[0014] Step 2: Image fusion: The infrared and visible light source images are decomposed into a basic part and a detailed part. This method fuses the decomposed images of the basic part to obtain a pre-fused basic layer F. b; For the details, first extract the depth features, that is, the component features and possible SF6 infrared features; then obtain the weight map through the multi-layer fusion strategy, and then use these weight maps and detail content to reconstruct the fused pre-fusion detail layer F d ; Finally, the pre-fused base layer F b and the pre-fused detail layer F d Fusion to obtain the optimized image F. The specific steps are as follows: Step 2.1, pre-fusion base layer construction: The base parts extracted from the source image contain common features and redundant information. In this method, a weighted average strategy is selected to fuse these base parts. The fused base layer is calculated by equation (4): (4); Where (x, y) represents the image intensity at , and The corresponding position in . and Respectively and In order to retain common features and reduce redundant information, in this method, we choose and .

[0015] Step 2.2: Pre-fusion detail layer construction: details separated from the source image and After processing, a new fusion is performed. This method uses a deep learning method (VGG-19 network) to extract the deep features of the detail content, and then obtains the weight map through a multi-layer fusion strategy. Finally, the pre-fused detail layer is obtained by reconstructing and fusion of these weight maps and detail content.

[0016] This method first obtains its deep features , represents the feature map of the nth detail content extracted by the kth layer, and m is the number of channels of the kth layer.

[0017] (5); In the formula, represents an activation layer in the VGG network, let , which means that four activation layers are selected from the VGG network as the source of deep features. , .

[0018] set up Represents the position in the feature map Department content, is an M-dimensional vector. The l1 norm of can be the activity level measure of the source detail content, and the initial activity level map is obtained by the following equation: (6); Based on the initial activity level map, we use a block-based average operator to calculate and obtain a robust activity level map.

[0019] (7); Where q determines the block size. This method chooses q = 1. If r is larger, it will be more robust to the registration errors of the last method, but some details may be lost.

[0020] Getting an activity level graph Finally, the soft maximum operator is used to calculate the initial weight map.

[0021] (8); Where N=2, Represents the initial weight map value in the range [0,1].

[0022] After obtaining each initial weight map Afterwards, we use an upsampling operator to resize the weight map to a size equal to the input detail content size.

[0023] (9); Through the above formula, four pairs of weight mapping sets can be obtained. Based on each pair of mapping sets, the following equation is used to calculate the alternative fusion detail layer.

[0024] (10); Finally, the maximum value is selected from the four initial fused detail contents at each position to obtain the fused detail content. .

[0025] Finally, the fused pre-fused base layer is used by this method and fused pre-fused detail layers Add them together to reconstruct the final fused image.

[0026] Embodiment 1: In order to verify the effectiveness of the proposed SF6 gas leakage detection method based on deep learning and visible light and infrared image fusion, this method was verified in the simulation software MATLAB R2018a.

[0027] The experimental object selected is the insulated power equipment with known SF6 leakage points. The source image is obtained by using the detection equipment, so that the fusion ratio of infrared and visible light images is 1:1. The proposed GIS equipment SF6 leakage identification method using visible light and infrared light image fusion is used for the experiment. 10 pairs of images are selected as the fusion data set, and one pair is selected as the image fusion result. The fusion result is shown in Figure 3 shown.

[0028] In image quality assessment, qualitative evaluation methods are very important, which are based on the human visual system to evaluate the quality of the fused image. The obtained image has obvious traces of gas leakage, strong contrast, the target is easy to separate from the scene, and many source image details are retained, which is conducive to identifying the type of equipment leaking gas.

[0029] It is very important to evaluate image quality. Different applications have different parameter requirements. In order to ensure the rigor of the evaluation, this method selects the following evaluation indicators.

[0030] In many image processing tasks, images are often represented by their features, and most of the information related to the quality of the fused image includes certain features, such as edges, details, and contrast. The feature mutual information (FMI) quality index is adopted to measure the amount of feature information transferred from the source image to the fused image. The definition of the FMI metric is as follows: (11); In the formula, H i (A), H i (B) and H i (F) is the entropy of the windows corresponding to the three images. i (A;F) and P i (B;F) represents the source image. The larger its value is, the more feature information is transferred from the source image to the fused image, which means that the fused image has more details.

[0031] The peak signal-to-noise ratio (PSNR) metric is the ratio of the peak power to the noise power in the fused image, and therefore reflects the distortion in the fusion process. The PSNR metric is defined as follows: (12); Where r represents the peak value of the fused image. The larger the PSNR, the closer the fused image is to the source image, and the smaller the distortion produced by the fusion method.

[0032] The root mean square error (RMSE) metric is similar to the MSE metric and is defined as follows: ; In the formula, RMSE AF and RMSEBF Represents the difference between the fused image and the infrared / visible image. A small RMSE metric indicates that the fused image has less error and distortion, and the contrast is greater.

[0033] The structural similarity of images (SSIM) is a metric used to compare the similarity of two images. The goal of SSIM is to simulate the human eye's perception of images, rather than simply comparing pixel values. It takes into account three aspects: brightness, contrast, and structure. The definition of SSIM includes three main components: Brightness represents the overall brightness level of the image, that is, grayscale information. In SSIM, brightness comparison is measured using the brightness average, considering the difference between the brightness averages of the two images. Contrast represents the degree of variation in pixel intensity in the image. In SSIM, contrast comparison is measured using the standard deviation, that is, the breadth of the pixel value distribution. Contrast comparison reflects the texture and detail information of the image. Structure represents the spatial relationship between pixels in the image. In SSIM, structural comparison is measured using covariance, that is, the correlation between pixel values. This part focuses on the structural information of the image, which helps capture the subtle texture of the image with contrast comparison. SSIM is defined as follows: (15); In the formula, SSIM(·) represents the structural similarity between the source image P and the fused image F. It reflects the ability to preserve structural information. The larger the value, the better the fusion performance.

[0034] The quantitative comparison of FMI, PSNR, RMSE and SSIM of the fused images obtained by the classical method and the proposed method is as follows Figure 5 shown.

[0035] from Figure 5 It can be seen that for the peak signal-to-noise ratio and root mean square error indicators, this method better preserves the structural information, produces less distortion, can better highlight the target information, and is more efficient. Combining the above qualitative and quantitative analysis, this method can make the target in the infrared and low-light images stand out, enrich the information of the infrared image, improve the image quality, and improve the analysis of SF6 gas leakage detection. It proves that this method has better practicality and convenience.

[0036] The above embodiments are only preferred technical solutions of the present invention and should not be regarded as limiting the present invention. The protection scope of the present invention shall be the technical solutions recorded in the claims, including equivalent replacement solutions of the technical features in the technical solutions recorded in the claims. That is, equivalent replacement improvements within this scope are also within the protection scope of the present invention.

Claims

1. A GIS equipment SF6 leakage identification method based on component and SF6 infrared feature image fusion, characterized in that The following steps are involved: Step 1: Image separation: Step 1.1, construct grayscale image: The source image of SF6 gas leakage is obtained as a color image under visible light and infrared light. Before image separation, the color image is converted into a grayscale image; Step 1.2, decomposing the source image: decomposing the source image into a base layer image and an original image; once the base layer image is obtained, the detail layer image, i.e., the detail layer image containing component features and possible SF6 infrared features, can be obtained by subtracting the basic part from the source image; Step 2: Image fusion: Step 2.1, pre-fusion base layer construction: The base parts extracted from the source image contain common features and redundant information, and a weighted average strategy is selected to fuse these base parts; Step 2.2: Pre-fusion detail layer construction: The detail parts separated from the source image are processed and then newly fused.

2. The GIS equipment SF6 leakage identification method based on component and SF6 infrared characteristic image fusion according to claim 1 is characterized by: In step 1.1, the color image is stored in the computer in the form of a three-dimensional matrix of M×N×3, where M represents the number of rows and N represents the number of columns of the image. After processing, it is converted into a two-dimensional matrix of M×N for storage; the conversion method and its calculation formula are as follows: （1） Where α1, α2 and α3 represent the weights of pixels in PR, PG and PB, respectively.

3. The GIS equipment SF6 leakage identification method based on component and SF6 infrared characteristic image fusion according to claim 2 is characterized in that: In step 1.2, assume that the visible light image and infrared image of the obtained source image are P1 and P2 respectively; decompose the source image into two scale representations, where the base layer image is obtained by solving the following formula: (2)； In the formula, is the base layer image, is the original image, is the convolution operation, G is the convolution kernel; is the detail layer image; once the base layer image is obtained, the detail layer image can be obtained by subtracting the basic part from the source image; (3)； Source image decomposition aims to separate each source image into a base layer image containing the basic frame and a detail layer image containing small-scale details; The source images are denoted as P1 and P2; the base layer image of each source image is obtained by solving equations (2) and (3) and detail layer images , where n∈{1,2}.

4. The GIS equipment SF6 leakage identification method based on component and SF6 infrared characteristic image fusion according to claim 3 is characterized in that: In step 2.1, the base parts extracted from the source image contain common features and redundant information; a weighted average strategy is selected to fuse the contents of these base parts; the fused base layer is calculated by equation (4): (4)； Where (x, y) represents the image intensity at , and The corresponding position in ; and Respectively and The weight value of the pixel in .

5. The GIS equipment SF6 leakage identification method based on component and SF6 infrared characteristic image fusion according to claim 4 is characterized in that: In step 2.2, first obtain its deep features , represents the feature map of the nth detail content extracted by the kth layer, and m is the number of channels of the kth layer; (5)； In the formula, represents an activation layer in the VGG network, let , indicating that four activation layers are selected from the VGG network as the source of deep features; , ; set up Represents the position in the feature map Department content, is an M-dimensional vector; The l1 norm of can be the activity level measure of the source detail content, and the initial activity level map is obtained by the following equation: (6)； Based on the initial activity level map, we use the block-based average operator to calculate and obtain a robust activity level map; (7)； In the formula, q determines the block size; Getting an activity level graph After that, the soft maximum operator is used to calculate the initial weight map: (8)； Where N=2, Represents the initial weight mapping value in the range [0,1]; After obtaining each initial weight map Afterwards, the weight map is resized to the size of the input detail content using an upsampling operator: (9)； Through the above formula, four pairs of weight mapping sets can be obtained; based on each pair of mapping sets, the following equation is used to calculate the alternative fusion detail layer: (10)； Finally, the maximum value is selected from the four initial fused detail contents at each position to obtain the fused detail content. ; Using a fused pre-fused base layer and fused pre-fused detail layers Add them together to reconstruct the final fused image.

Images