High-noise resistance wave trapper thermal fault positioning method based on thermal region segmentation

Abstract

The invention discloses a high-noise resistance wave trapper thermal fault positioning method based on thermal region segmentation. The method comprises the following steps: 1, acquiring an infrared image; 2, removing infrared image noise by using a column weighting and hierarchical filtering algorithm and an SVD (Singular Value Decomposition) and information entropy linear fitting algorithm in sequence; 3, segmenting the infrared image by using a gradient weight region growing method; 4, obtaining an effective extreme point set according to the segmented infrared image; and 5, performing clustering processing according to the effective extreme point set, finding out a thermal anomaly region of the wave trapper according to a threshold condition, and drawing a mask at a corresponding position in the infrared image so as to obtain a heating anomaly positioning image. According to the method, high-quality noise removal and accurate segmentation can be realized in the image processing process, so that the abnormal heating identification precision and speed of the wave trapper can be improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of power equipment fault detection, and specifically relates to a method for thermal fault location of a high-noise-resistant arrester based on thermal region segmentation. Background Art

[0002] The normal operation of power equipment is crucial for industrial production and residents' lives. Compared with the defects or faults in the appearance of instruments, the abnormal heating of power equipment is not only difficult to directly detect but also affects power supply in a short time. Therefore, the real-time monitoring of the heating faults of power equipment has practical significance. Transformers, circuit breakers, power capacitors, and arresters are common power equipment in substations that are prone to abnormal heating. Among them, compared with equipment such as transformers and circuit breakers placed on the ground, it is difficult to judge the heating situation of the arresters placed in the air through direct contact manual detection. Therefore, designing a non-contact automatic judgment method for the heating situation can meet the needs of real-time monitoring of power equipment.

[0003] Many scholars have carried out research on the detection of abnormal heating of power equipment:

[0004] Ge Xinyuan constructed a hidden Markov tree model for wavelet coefficients in the paper "Research on Denoising of Insulator Infrared Images Based on Wavelet Domain Hidden Markov Tree Model and Importance Correction", obtained the noise variance through the Donoho noise estimation method, and removed the noise using the empirical Bayes formula; Quan Pan proposed to estimate the noise variance using the spatial domain correlation estimation method in the paper "Two Denoising Methods by Wavelet Transform"

[0005] Xiao Xiaohui et al. used the Roberts operator to complete the edge contour detection of the infrared images of power equipment in the paper "Edge Detection of Infrared Images of High-Voltage Transmission Lines", realizing the segmentation of equipment patterns; Patent CN112085037A combines the particle swarm optimization algorithm PSO and the Otsu global threshold algorithm to separate the background and the power equipment area, and then analyzes and processes the patterns of the equipment area.

[0006] Xiao Yi et al. performed probability density function fitting on the temperature histogram obtained from the infrared image in the paper "Thermal Infrared Image Recognition Method for Substation High-Voltage Equipment Faults Based on Temperature Probability Density", and used the extreme points where the probability density function conforms to the screening criterion as clustering temperature points and used the clustering algorithm for thermal region segmentation, but whether the region is abnormally heated still needs to be judged manually; Yin Yang used the SURF algorithm for feature point matching to locate the thermal abnormal area in the paper "Research on Substation Equipment Identification and Thermal State Monitoring System Based on Infrared Images"

[0007] Surface temperature method, similar comparison method, relative temperature difference method, etc. are commonly used methods for judging abnormal heating of power equipment. Zhang Wenfeng et al. proposed a judgment method combining surface temperature and relative temperature difference in the paper "Intelligent Diagnosis Technology for Heating Defects of Transmission Lines Based on UAV Infrared Video" for detecting heating faults of power equipment.

[0008] Although in recent years, typical power equipment heating fault diagnosis methods based on machine vision have been relatively fully studied in aspects such as equipment recognition, infrared image processing, and thermal fault judgment methods, there are still the following deficiencies:

[0009] 1. Most methods do not consider the strip noise caused by camera hardware. Common arresters have strip or lattice-shaped casings, and the strip noise will cover the image contour of the arrester, thus seriously interfering with the image region segmentation of the arrester and the positioning of abnormal heating points. At the same time, the current general filtering operators will lose image details when removing image noise, resulting in image blurring or even distortion.

[0010] 2. In the estimation of infrared image noise energy in the wavelet domain, the mainstream Donoho estimation method assumes that after wavelet decomposition of the noise image, there is very little original image information in the diagonal wavelet coefficients of the last layer, that is, it defaults that the original image contains high noise. However, in low-noise images, the noise estimation obtained by the Donoho method has a large deviation. Under the current working conditions, the infrared noise caused by optical factors is already weak, and the arrester image with a striped casing contains rich detail information. The Donoho method is prone to regarding edge detail information as noise components under low-noise conditions, resulting in image distortion.

[0011] 3. In the segmentation of the equipment image area, compared with other large power equipment stored independently, there are transmission lines, insulators, hollow brackets and other equipment around the arrester placed in the air. The threshold segmentation method in the existing solutions is extremely vulnerable to the interference of environmental images, thus wrongly selecting the connected domain. At the same time, the accuracy of the region growing algorithm depends on the selection of seed points, but the current improved region growing algorithm still uses color similarity as the criterion, lacking geometric configuration information of the target object as a further screening criterion.

[0012] 4. In the positioning of the abnormal thermal state area, since the abnormally heated part accounts for a small proportion in the infrared image, a small number of pixels are difficult to have sufficient feature points, limiting the accuracy of feature point matching algorithms such as SURF. Moreover, the neighborhood of the abnormally heated part will also show a temperature rise, making the region boundary blurred, resulting in a large deviation in the similarity comparison of the region template matching method; at the same time, although the clustering method based on feature region matching and temperature probability statistics has simple steps, it usually has difficulty in distinguishing background points and equipment area points in infrared images, thus introducing a large number of invalid samplings, resulting in a large number of iterations and low efficiency.

[0013] 5. In the abnormal diagnosis of hot areas, for current-induced heating problems, the common approach is to combine surface temperature with relative temperature difference for judgment to determine the abnormal heating threshold. However, in reality, current-induced heating and voltage-induced heating often coexist. Since the heat generation of voltage-induced heating equipment is small and internal defects are usually enclosed, the surface temperature cannot reflect the actual temperature of the defect point. Therefore, other criteria are needed to determine the abnormal heating threshold. Summary of the Invention

[0014] The present invention is proposed to solve the above-mentioned deficiencies of the prior art, and provides a high-noise-resistant arrester thermal fault location method based on thermal area segmentation, aiming to achieve high-quality denoising and accurate segmentation in the image processing process, so as to improve the accuracy and speed of abnormal heating identification of arresters.

[0015] To achieve the above object, the present invention adopts the following technical solutions:

[0016] A high-noise-resistant arrester thermal fault location method based on thermal area segmentation is characterized by performing according to the following steps:

[0017] Step 1: Collect an infrared image A of the arrester with a dimension of W×H 1 , where W and H respectively represent the width and height of the infrared image A 1 ;

[0018] Step 2: Denoise A 1 to obtain the denoised infrared image A 4 ;

[0019] Step 2.1: After converting A 1 into a grayscale image A 2 of the same scale, use the column weighted and hierarchical filtering algorithm to process A 2 to obtain an infrared image A 3 without stripe noise;

[0020] Step 2.2: After performing a logarithmic transformation on A 3 , then use the Coiflet wavelet basis to perform N-layer wavelet decomposition on A 3 to obtain a wavelet image set {G i |1≤i≤N}, where G i represents the wavelet image obtained by the i-th layer of decomposition;

[0021] Step 2.3: Use the algorithm based on SVD decomposition and information entropy linear fitting to process {G i |1≤i≤N} to obtain an infrared image A 4 without stripe noise and speckle noise;

[0022] Step 3: Use the gradient-weighted region growing method to segment the infrared image A 4 , and obtain the segmented infrared image A 5 ;

[0023] Step 4: According to the gray histogram of the segmented infrared image A 5 , obtain the gray probability density function, and thus use the gray probability density function to obtain the average gray value T of the arrester region in A 1 , and further obtain the set of effective extreme points of the arrester region in the three-dimensional gray distribution map corresponding to A device ; 5

[0024] Step 5: Take each coordinate in the set of effective extreme points as the clustering center, and use the KNN algorithm to cluster the pixel points in A 5 , and after obtaining the clustering information matrix, find the thermal anomaly region of the arrester from the clustering information matrix according to the threshold condition, and draw a mask at the corresponding position in A 1 to obtain the thermal anomaly localization image A of the arrester 6 .

[0025] The feature of a high-noise-resistant arrester thermal fault localization method based on thermal region segmentation according to the present invention also lies in that the column weighting and hierarchical filtering algorithm in step 2.1 includes the following steps:

[0026] Step 2.1.1: Use the gradient operator to process the gray image A 2 to obtain the horizontal gradient map Grad;

[0027] Step 2.1.2: Define the column weighting filter window:

[0028] Define a filter window with a dimension of (2k + 1)×(2k + 1), where 2k + 1 is the dimension of the filter window and k is a positive integer;

[0029] Use formula (1) to obtain the pixel value S(i, j) at the center position (i, j) of the filter window in A 2 , and set the boundary constraint condition of the filter window as: , , and thus slide on the gray image A 2 in the order from left to right and from top to bottom to obtain the sliding window image B containing low-frequency information 1 ;

[0030] (1)

[0031] In formula (1), represents A 2 ​The average pixel value of the x-th column in, and , represents A 2 The pixel value at the coordinate (x, y) in; represents the correction function for A x , and , represents a manually determined regulation factor, represents A 2 The average pixel value of the (x + 1)-th column in, sgn(·) represents the sign function, represents the average pixel value of the x-th column in Grad, and , represents the pixel value at the coordinate (x, y) in Grad; represents A 2 The column weight of the x-th column in, and , σ is the set variance of the column weights;

[0032] Step 2.1.3. Subtract B 2 from A 1 to obtain the difference image B 2 , and perform high-pass filtering on B 2 in the column direction to obtain the high-frequency information image B 3 ;

[0033] Step 2.1.4. Process B 3 using Equation (2) to obtain the corrected image B 4 :

[0034] (2)

[0035] In Equation (2), B 4 (x, y) and B 3 (x, y) respectively represent the pixel values at the coordinate (x, y) in B 4 and B 3 ; represents the average value of the pixel points in B 3 ; λ is a manually determined denoising factor;

[0036] Step 2.1.5. Add B 1 and B 4 to obtain the true grayscale image B 5 , and convert B 5 to the same color space as A 1 to obtain the infrared image A 3 without stripe noise.

[0037] Further, the SVD decomposition and information entropy linear fitting algorithm in step 2.3 includes the following steps:

[0038] Step 2.3.1: Perform SVD decomposition on each wavelet image in the wavelet image set {G i | 1 ≤ i ≤ N} in the horizontal sub-band and vertical sub-band directions to obtain a set of non-zero singular value sequences {(Q i , P i )| 1 ≤ i ≤ N}, where Q i is the non-zero singular value sequence of the i-th wavelet image G i in the horizontal sub-band, and Q i ={Q i,s | 1 ≤ s ≤S i}, Q i,s represents the s-th horizontal non-zero singular value in Q i , and S i represents the total number of horizontal non-zero singular values in Q i ; P i is the non-zero singular value sequence of the i-th wavelet image G i in the vertical sub-band, and P i ={P i,t | 1 ≤t ≤T i}, P i,t represents the t-th vertical non-zero singular value in Q i , and T i represents the total number of vertical non-zero singular values in Q i ;

[0039] Step 2.3.2: Traverse {(Q i ,P i ) | 1 ≤ i≤ N}, and sequentially remove the non-zero singular values in Q i and P i that are less than the median to obtain a set of non-zero singular value sequences after removal {(Q i ’,P i ’) | 1≤i≤N}; where Q i ’is the non-zero singular value sequence of the i-th wavelet image G i after removal in the horizontal sub-band; P i ’is the non-zero singular value sequence of the i-th wavelet image G i after removal in the vertical sub-band;

[0040] Step 2.3.3: Traverse {(Q i ’,P i ’) | 1≤i≤N}, and calculate the average value Q i ’of Q i,ave’ and P i The average value of P i,ave ’, and respectively select Q i ’ that is greater than Q i,ave ’s non-zero singular values and P i ’ that is greater than P i,ave ’s non-zero singular values for reconstruction, and obtain the reconstructed wavelet image set {G i ’ | 1 ≤ i ≤ N}; where, G i ’ represents the wavelet image after the i-th layer of reconstruction;

[0041] Step 2.3.4, calculate the information entropy H i ’ of the diagonal sub-band wavelet coefficients of G i,1 , and thus obtain the transformed information entropy H i,2 using Equation (3):

[0042] (3)

[0043] Step 2.3.5, perform polynomial fitting on H i,2 based on prior information, and calculate the noise variance σ i of the diagonal sub-band of G i :

[0044] Step 2.3.6: Use the empirical Bayesian formula and σ i to estimate the true wavelet coefficients of {G i ’ | 1 ≤ i ≤ N}, and thus obtain the true wavelet image set {G i ’’|1≤i≤N}, where, G i ’’ represents the i-th layer true wavelet image;

[0045] Step 2.3.7, perform exponential transformation and inverse wavelet transformation on each layer of the true wavelet image in {G i ’’|1≤i≤N} in sequence, and then perform image reconstruction to obtain the infrared image A 4 without stripe noise and speckle noise.

[0046] Furthermore, the gradient weight region growing method in Step 3 includes the following steps:

[0047] Step 3.1, after converting A 4 to a grayscale image A 4 ’ of the same scale, perform horizontal and vertical gradient detections on the grayscale image A 4 ’ respectively, and correspondingly obtain the horizontal gradient map Grad x and the vertical gradient map Grad y ;

[0048] Step 3.2, A4 After converting to the HSV color space, the watershed algorithm is used to pre-segment the image after the HSV color space conversion to obtain the first sub-region set , where represents the u-th first sub-region, and the set of adjacent sub-regions of is denoted as which represents the v-th adjacent sub-region of and represents the total number of adjacent sub-regions of

[0049] Step 3.3: Initialize u = 1;

[0050] Step 3.4: Calculate the root-mean-square difference of the saturation component simS between and u the root-mean-square difference of the hue component simH u , and calculate the root-mean-square difference of the gradient covariance simG between and u using Equation (4);

[0051] (4)

[0052] In Equation (6), and respectively represent the mean horizontal gradient and the mean vertical gradient of and respectively represent the mean horizontal gradient and the mean vertical gradient of the sub-region and all its adjacent sub-regions;

[0053] Step 3.5: Calculate the similarity function f of u using Equation (5); if f u is less than the threshold , then mark as a seed region and go to Step 3.6; otherwise, directly go to Step 3.6;

[0054] (5)

[0055] In Equation (5), , , are three weight coefficients respectively;

[0056] Step 3.6: If u is less than sub 1 , then assign u + 1 to u and return to Step 3.4 to execute sequentially; otherwise, it means the seed region set is obtained;

[0057] Step 3.7: Select all seed regions from the seed region set for growth according to the seed growth rule to obtain a second sub-region set;

[0058] Step 3.8: Perform rectangular fitting on all sub-regions in the second sub-region set, and remove the sub-regions whose aspect ratio of the fitted rectangle is greater than the threshold of the sub-regions. Sort the remaining sub-regions in descending order according to the distance from the geometric center to the A 4 center to obtain a third sub-region set. At the same time, sort the remaining sub-regions in descending order according to their own region area to obtain a fourth sub-region set; Select the sub-region with the smallest sum of indices in the third sub-region set and the fourth sub-region set as the segmented infrared image A 5 .

[0059] Further, the said step 4 includes the following steps:

[0060] Step 4.1: Grayscale A 5 , and draw a grayscale histogram according to the grayscale values of the pixel points, so as to obtain the grayscale probability density function F(t) of A 5 by using the kernel estimation method, where t represents the grayscale value, and t ∈ (T min , T max ); Among them, T min , T max are respectively the minimum value and the maximum value in the grayscale histogram;

[0061] Step 4.2: Use the otsu algorithm to obtain the segmentation threshold T of F(t), and regard the pixel points in A 5 with pixel values less than T as background points, and the pixel points with pixel values greater than T as blocker region points; Thus, the average grayscale value T device of the blocker region and the average grayscale value T env of the background region are obtained by using equations (6) and (7) respectively;

[0062] (6)

[0063] (7)

[0064] Step 4.3: Draw a three-dimensional grayscale distribution map of A 2 on the rectangular coordinate system. Among them, the x-axis and y-axis of the rectangular coordinate system represent the abscissa and ordinate of the image pixel points, and the z-axis represents the corresponding grayscale value of the pixel points;

[0065] Define the plane z 0 , and initialize it as z 0 = T device ;

[0066] Step 4.4: Assign to z 0 . If there is a point tangent to z = z 0 in the three-dimensional gray distribution map, store the x and y axis coordinates of the tangent point in the extreme point set, and go to Step 4.5; otherwise, directly go to Step 4.5; where is a set parameter;

[0067] Step 4.5: Determine whether z 0 is less than T max . If so, go to Step 4.4; otherwise, output the coordinate set of valid extreme points {(x m , y m )|1 ≤ m ≤ E max}; where (x m , y m ) represents the coordinates of the m-th valid extreme point, and E max represents the total number of extreme points.

[0068] Furthermore, Step 5 includes the following steps:

[0069] Step 5.1: Use each coordinate in the valid extreme point set as a clustering center, and perform clustering on the points in the arrester area using the KNN algorithm to obtain the clustering information matrix , where A max represents the total number of clustering regions, S e represents the area of the e-th clustering region, and T e represents the average gray value of the e-th clustering region; represents the relative temperature difference of the e-th clustering region, and ; W e represents the weight of the e-th clustering region, and ;

[0070] Step 5.2: Calculate the threshold T th using Equation (8):

[0071] (8)

[0072] In Equation (8), is the correction factor;

[0073] Step 5.3: Initialize e = 1;

[0074] Step 5.4: If is greater than the threshold T th , then according to the position of the e-th clustering region in the infrared image A 1 , in the infrared image A 1Draw a mask at the corresponding position in [the image], and go to step 5.5; otherwise, directly go to step 5.5;

[0075] Step 5.5: If e is less than A max , then assign e + 1 to e, and then go to step 5.4 to execute sequentially; otherwise, it means that an infrared image with a mask is obtained and used as the thermal anomaly localization image A 6 Output.

[0076] An electronic device according to the present invention includes a memory and a processor, characterized in that the memory is used to store a program for supporting the processor to execute the high-noise-resistant arrester thermal fault localization method, and the processor is configured to execute the program stored in the memory.

[0077] A computer-readable storage medium according to the present invention, characterized in that a computer program is stored on the computer-readable storage medium, and the computer program executes the steps of the high-noise-resistant arrester thermal fault localization method when being run by a processor.

[0078] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0079] 1. Since the stripe noise in the image is in a longitudinal form and the pixel values of the noise pixel points in the same column are similar, the present invention designs a new filtering operator, and using the average value of the pixel values of each column of the grayscale image as the processing unit better fits the characteristics of the stripe noise; since the stripe noise intensities of different columns in the image are different and the differences from the adjacent real patterns are also different, the new filtering operator combines the horizontal gradient information, and according to the change rate of the pixel values in the horizontal direction and the noise intensity itself, numerically compensates the noise pixel points, which can achieve an adaptive smoothing effect and alleviate the blurring effect caused by over-smoothing at weak noise areas; based on the pixel value continuity of the real pattern, the new filtering operator sets weights according to the Gaussian distribution for the horizontal distances from different columns to the center of the image, strengthens the influence of the nearby pixel information on the center of the operator, and suppresses the image distortion caused by the distant information.

[0080] 2. The column weighting and hierarchical filtering algorithm proposed by the present invention, after smoothing the original image to obtain a smoothed image, subtracts the original image from the smoothed image and performs horizontal high-pass filtering to obtain a high-frequency image containing a large amount of noise. At the same time, the high-frequency image is filtered again according to the horizontal direction gradient information to obtain a smoothed high-frequency image, and the smoothed high-frequency image and the smoothed image are added together to obtain a denoised image. Compared with single filtering, smoothing the noisy high-frequency image can alleviate the information loss caused by over-smoothing, and the denoised image obtained according to the method of the present invention further alleviates the blurring effect caused by smoothing.

[0081] 3. The SVD decomposition and information entropy linear fitting algorithm proposed by the present invention uses an adaptive screening method to remove the smaller singular values in the horizontal sub-band and vertical sub-band obtained by the SVD decomposition of the wavelet image, and can retain the larger singular values to obtain the real pattern. At the same time, when the noise intensity decreases, compared with the traditional Donoho estimation method, the linear relationship between the entropy value of the noisy image in the wavelet sub-band and the noise variance is used to estimate the noise energy variance of the diagonal sub-band of the wavelet image, which is closer to the real value and is not easily interfered by image details.

[0082] 4. The present invention uses an improved seed region growing method to extract the device pattern region, and performs image pre-segmentation processing through the watershed algorithm to obtain a large number of sub-regions. In view of the hollow appearance characteristics of the arrester device, based on the regional similarity and Euclidean distance criterion, gradient information is introduced as a new similarity criterion, so that more appropriate seed regions can be selected, avoiding the generation of wrong regions due to improper seed selection. And the distance from the region to the center of the image and its own area are used as screening criteria to further screen the real device pattern, thus alleviating the interference of external background devices in the image.

[0083] 5. The present invention uses the otsu algorithm to obtain the pixel value ranges of the background points and device region points from the gray histogram, so as to further distinguish the background region and device region in the image, improving the segmentation accuracy of the device hot region. And, based on the known condition that the pixel gray information can reflect the temperature information, by constructing a three-dimensional gray distribution map, the average gray value of the background region is used as the starting point to find the extreme point in the high pixel value region and used as the clustering center, so as to determine the high temperature points in the two-dimensional image, thus reducing the invalid sampling and improving the clustering speed.

[0084] 6. Under the voltage-induced heating problem, based on the characteristic of the small area of the abnormal heating region, the present invention introduces the clustering region area as an influencing factor to determine the abnormal heating threshold, so as to reduce the false positive rate of thermal anomaly recognition when the heating phenomenon is not obvious. BRIEF DESCRIPTION OF THE DRAWINGS

[0085] Figure 1 is the flow chart of the steps of the present invention;

[0086] Figure 2 is the schematic diagram of wavelet decomposition in the present invention;

[0087] Figure 3 is the comparison chart of the noise standard deviation estimation results of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0088] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. To make the purpose, features, and advantages of the present invention more obvious and understandable, the present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0089] In this embodiment, a method for locating the thermal fault of a high-noise-resistant arrester based on thermal region segmentation is as Figure 1 shown and is carried out according to the following steps:

[0090] Step 1: Collect the infrared image A of the arrester with a dimension of W×H 1 , where W and H respectively represent the width and height of the infrared image A 1 ; in this embodiment, W is 1024 and H is 1024.

[0091] Step 2: Perform denoising processing on A 1 to obtain the denoised infrared image A 4 ;

[0092] Step 2.1: After converting A 1 into a grayscale image A of the same scale 2 , use the column weighted and hierarchical filtering algorithm to process A 2 to obtain an infrared image A 3 without stripe noise.

[0093] Step 2.1.1: Use the gradient operator to process the grayscale image A 2 to obtain the horizontal gradient map Grad;

[0094] Step 2.1.2: Define the column weighted filtering window:

[0095] Define a filtering window with a dimension of (2k + 1)×(2k + 1), where 2k + 1 is the dimension of the filtering window. In this embodiment, k takes the positive integer 3.

[0096] Construct a loop program to implement the traversal of pixel points, and use Equation (1) to obtain the pixel value S(i, j) at the center position (i, j) of the filtering window in A 2 , and set the boundary constraint conditions of the filtering window as: , , so that according to the boundary constraint conditions, slide on the grayscale image A 2 from left to right and from top to bottom to obtain the sliding window image B 1 ;

[0097] (1)

[0098] In formula (1), represents the average pixel value of the x-th column in A 2 , and , represents the pixel value at the coordinate (x, y) in A 2 ; represents the correction function with respect to A x , and , represents a manually determined regulation factor. In this implementation, the value is 0.2, represents the average pixel value of the (x + 1)-th column in A 2 . sgn(·) represents the sign function, represents the average pixel value of the x-th column in Grad, and , represents the pixel value at the coordinate (x, y) in Grad; f(x) represents the column weight of the x-th column in A 2 , and , σ is the variance of the manually determined column weight. In this embodiment, σ takes the value of 1.2.

[0099] Step 2.1.3. Subtract B 2 from A 1 to obtain the difference image B 2 , and perform high-pass filtering on B 2 in the column direction to obtain the high-frequency information image B 3 ;

[0100] Step 2.1.4. Process B 3 using formula (2) to obtain the corrected image B 4 :

[0101] (2)

[0102] In formula (2), B 4 (x, y) and B 3 (x, y) respectively represent the pixel values at the coordinate (x, y) in B 4 and B 3 ; represents the average value of the pixel points in B 3 ; λ is a manually determined denoising factor; in this embodiment, λ takes the value of 0.15.

[0103] Step 2.1.5. Add B 1 and B 4 to obtain the true grayscale image B 5 , and convert B 5 to be the same as A1 the same color space to obtain an infrared image A without stripe noise 3 .

[0104] Step 2.2. After performing a logarithmic transformation on A 3 and then using the Coiflet wavelet basis to perform N-level wavelet decomposition on A 3 , as Figure 2 shown, in this embodiment, 3-level wavelet decomposition is performed and N is taken as 3; and Figure 2 in LH1, LH2, LH3 are horizontal sub-bands, HL1, HL2, HL3 are vertical sub-bands, and HH1, HH2, HH3 are diagonal sub-bands; to obtain a wavelet image set {G i | 1 ≤ i ≤ N}, where G i represents the wavelet image obtained by the i-th level of decomposition.

[0105] Step 2.3. Use the algorithm based on SVD decomposition and information entropy linear fitting to process {G i |1≤i≤N} to obtain an infrared image A without stripe noise and speckle noise 4 ;

[0106] Step 2.3.1. Respectively perform SVD decomposition on each wavelet image in the wavelet image set {G i | 1 ≤ i ≤ N} in the horizontal sub-band and vertical sub-band directions to obtain a set of non-zero singular value sequences {(Q i , P i )| 1 ≤ i ≤ N}, where Q i is the non-zero singular value sequence of the i-th wavelet image G i in the horizontal sub-band, and Q i ={Q i,s | 1 ≤ s ≤S i}, Q i,s represents the s-th horizontal non-zero singular value in Q i , and S i represents the total number of horizontal non-zero singular values in Q i ; P i is the non-zero singular value sequence of the i-th wavelet image G i in the vertical sub-band, and P i ={P i,t | 1 ≤t ≤T i}, P i,t represents the t-th vertical non-zero singular value in Q i , and T i represents the total number of vertical non-zero singular values in Q i .

[0107] Step 2.3.2, traverse {(Q i , P i ) | 1 ≤ i ≤ N}, and sequentially remove the non - zero singular values less than the median in Q i and P i to obtain the set of non - zero singular value sequences after removal {(Q i ’, P i ’) | 1 ≤ i ≤ N}; where, Q i ’ is the non - zero singular value sequence after removal in the horizontal sub - band of the i - th wavelet image G i ; P i ’ is the non - zero singular value sequence after removal in the vertical sub - band of the i - th wavelet image G i .

[0108] Step 2.3.3, traverse {(Q i ’, P i ’) | 1 ≤ i ≤ N}, calculate the average value Q i ’ of Q i,ave ’ and the average value P i ’ of P i,ave ’ respectively, and select the non - zero singular values greater than Q i ’ in Q i,ave ’ and the non - zero singular values greater than P i ’ in P i,ave ’ for reconstruction, then obtain the set of reconstructed wavelet images {G i ’ | 1 ≤ i ≤ N}; where, G i ’ represents the i - th layer reconstructed wavelet image.

[0109] Step 2.3.4, calculate the information entropy H i of the diagonal sub - band wavelet coefficients of G i,1 , and thus obtain the transformed information entropy H i,2 using Equation (3):

[0110] (3)

[0111] Step 2.3.5, perform polynomial fitting on H i,2 based on prior information, and calculate the noise variance σ i of the diagonal sub - band of G i : In this embodiment, through the data statistics of 240 images, the calculation formula of σ i is obtained as follows:

[0112]

[0113] Step 2.3.6: Use the empirical Bayes formula and σ iEstimate the true wavelet coefficients of {G i ’ | 1 ≤ i ≤ N}, so as to obtain the set of true wavelet images {G i ’’ | 1 ≤ i ≤ N}, where G i ’’ represents the true wavelet image of the i-th layer; in this embodiment, the Bayesian empirical formula used is as shown in Equation (3a):

[0114] (3a)

[0115] In Equation (3a), represents the j-th true wavelet coefficient in the diagonal subband of G i ’’, represents the j-th wavelet coefficient in the diagonal subband of G i ’, represents the variance observed for the j-th wavelet coefficient in the diagonal subband of G i ’.

[0116] As Figure 3 shown, is the noise standard deviation of the preset values during the simulation and verification of this embodiment, which are 15, 20, 25, 30, and 35 respectively, is the standard deviation estimated by this method, is the standard deviation obtained by the Donoho estimation method. After testing 500 images, compared with the traditional Donoho estimation method, the method used in this embodiment estimates the true wavelet coefficients more accurately. Especially when the noise variance is small, the effect of this method is significantly better than that of the Donoho estimation method.

[0117] Step 2.3.7. After sequentially performing exponential transformation and inverse wavelet transformation on each layer of the true wavelet image in {G i ’’ | 1 ≤ i ≤ N}, and then performing image reconstruction, an infrared image A 4 without stripe noise and speckle noise is obtained.

[0118] Step 3. Use the gradient-weighted region growing method to segment the infrared image A 4 , and obtain the segmented infrared image A 5 ;

[0119] Step 3.1. After converting A 4 into a grayscale image A 4 ’ of the same scale, perform horizontal and vertical gradient detections on the grayscale image A 4 ’ respectively, and correspondingly obtain the horizontal gradient map Grad x and the vertical gradient map Grad y ;

[0120] Step 3.2. Convert A4 After converting to the HSV color space, the watershed algorithm is used to pre-segment the image converted in the HSV color space to obtain the first set of sub-regions , where represents the u-th first sub-region, and the set of adjacent sub-regions of is denoted as which represents the v-th adjacent sub-region of and represents the total number of adjacent sub-regions of

[0121] Step 3.3: Initialize u = 1;

[0122] Step 3.4: First, calculate the root-mean-square difference simS of the saturation components between and . In this embodiment, simS u is obtained by the following formula (3b): u

[0123] (3b)

[0124] In formula (3b), represents the mean value of the saturation components of represents the mean value of the saturation components of all regions in and

[0125] Next, calculate the root-mean-square difference simH of the hue components u . In this embodiment, simH u is obtained by the following formula (3c):

[0126] (3c)

[0127] In formula (3c), represents the mean value of the hue components of represents the mean value of the hue components of all regions in

[0128] Finally, use formula (4) to calculate the root-mean-square difference simG of the gradient covariance between and ; u ;

[0129] (4)

[0130] In formula (4), and respectively represent The horizontal gradient mean value and the vertical gradient mean value of and respectively represent the horizontal gradient mean value and the vertical gradient mean value of the sub-region and all its adjacent sub-regions.

[0131] Step 3.5: Calculate the similarity function f of using Equation (5); if f u is less than the threshold u , then mark as a seed region and go to Step 3.6; otherwise, directly go to Step 3.6; (5)

[0132] In Equation (5),

[0133] are three weight coefficients respectively. In this embodiment, , , take 0.6, take 0.2, take 0.2, take 0.2.

[0134] Step 3.6: If u is less than sub 1 , then assign u + 1 to u and return to Step 3.4 to execute sequentially; otherwise, it means that the set of seed regions is obtained;

[0135] Step 3.7: According to the seed growth rule, select all seed regions from the set of seed regions for growth to obtain the second set of sub-regions.

[0136] Step 3.8: Perform rectangular fitting on all sub-regions in the second set of sub-regions, and remove the sub-regions with the aspect ratio of the rectangle after fitting greater than the threshold . Arrange the remaining sub-regions in descending order according to the distance from the geometric center to the center of A 4 to obtain the third set of sub-regions. At the same time, arrange the remaining sub-regions in descending order according to their own region areas to obtain the fourth set of sub-regions; select the sub-region with the smallest sum of indices in the third set of sub-regions and the fourth set of sub-regions as the segmented infrared image A 5 .

[0137] Step 4: According to the gray histogram of the segmented infrared image A 5 , obtain the gray probability density function, and thus use the gray probability density function to obtain the average gray value T 1 of the arrester region in A device , and further obtain the set of effective extreme points of the arrester region in the three-dimensional gray distribution map corresponding to A 5 ;

[0138] Step 4.1. Gray-scale A 5 , and draw a gray-scale histogram based on the gray-scale values of the pixels, so as to obtain the gray-scale probability density function F(t) of A 5 using the kernel estimation method, where t represents the gray-scale value and t ∈ (T min , T max ); among them, T min and T max are the minimum and maximum values in the gray-scale histogram respectively; in this embodiment, the kernel function used is the Gaussian distribution kernel function.

[0139] Step 4.2. Use the Otsu algorithm to obtain the segmentation threshold T of F(t), and use the pixels in A 5 with pixel values less than T as background points and the pixels with pixel values greater than T as blocker region points; thus, the average gray-scale value T device of the blocker region and the average gray-scale value T env of the background region are obtained using equations (6) and (7) respectively;

[0140] (6)

[0141] (7)

[0142] The principle of this step is that the number of abnormally heated pixels in the device region is small, so the number of pixels with high gray-scale values is small. Therefore, the gray-scale histogram can be approximately composed of two peaks, which respectively correspond to the peak of the background region and the peak of the device region. The heated region is ignored, and then the Otsu algorithm for processing binary classification problems can be introduced.

[0143] Step 4.3. Draw a three-dimensional gray-scale distribution map of A 2 in a rectangular coordinate system, where the x-axis and y-axis of the rectangular coordinate system represent the abscissa and ordinate of the image pixels, and the z-axis represents the gray-scale value corresponding to the pixels;

[0144] Define the plane z 0 , and initialize it to z 0 = T device ;

[0145] Step 4.4. Assign to z 0 . If there is a point tangent to z = z 0 in the three-dimensional gray-scale distribution map, store the x and y axis coordinates of the tangent point in the extreme point set, and go to step 4.5; otherwise, directly go to step 4.5; where is a set parameter;

[0146] Step 4.5. Judge z0 Is it less than T max , if yes, go to step 4.4; otherwise, output the coordinate set of valid extreme points \(\{(x m , y m )|1 \leq m \leq E max}\), where \((x m , y m ) represents the coordinates of the \(m\)-th valid extreme point, and \(E max represents the total number of extreme points.

[0147] Step 5: Take each coordinate in the set of valid extreme points as the clustering center, and use the KNN algorithm to cluster the pixel points in A 5 to obtain a clustering information matrix; then, according to the threshold condition, find the thermal anomaly area of the arrester from the clustering information matrix, and draw a mask at the corresponding position in A 1 to obtain the thermal anomaly localization image A 6 .

[0148] Step 5.1: Take each coordinate in the set of valid extreme points as the clustering center respectively, and use the KNN algorithm to cluster on the arrester area points to obtain a clustering information matrix:

[0149]

[0150] where \(A max represents the total number of clustering regions, \(S e represents the area of the \(e\)-th clustering region, \(T e represents the average gray value of the \(e\)-th clustering region; represents the relative gray difference of the \(e\)-th clustering region, and ; \(W e represents the weight of the \(e\)-th clustering region, and .

[0151] Step 5.2: Calculate the threshold \(T th using Equation (8). In Equation (8), is the correction factor;

[0152] (8)

[0153] In this embodiment, is taken as 7; at the place where the embodiment is executed, usually \(T device is taken as about 150, \(T env is taken as about 65; according to actual tests, the \(W e of the abnormal heating point is taken as about 0.02; generally speaking, when the gray value exceeds 210, it can be considered that the heating is abnormal.

[0154] Step 5.3: Initialize \(e = 1\) and start traversing the clustering information matrix from the first row;

[0155] Step 5.4: If is greater than the threshold \(T\) th , then draw a mask at the corresponding position in the infrared image \(A\) according to the position of the \(e\)-th clustering region in the infrared image \(A\) 1 , and transfer to Step 5.5; otherwise, directly transfer to Step 5.5; 1 Step 5.5: If \(e\) is less than \(A\)

[0156] , then assign \(e + 1\) to \(e\) and transfer to Step 5.4 to execute sequentially; otherwise, it means that the infrared image with the mask is obtained and output as the thermal anomaly localization image \(A\) max . 6 Output.

[0157] In this embodiment, an electronic device includes a memory and a processor. The memory is used to store a program that supports the processor to execute the above method, and the processor is configured to execute the program stored in the memory.

[0158] In this embodiment, a computer-readable storage medium stores a computer program, and when the computer program is run by a processor, it executes the steps of the above method.

Claims

1. A method for locating thermal faults of high noise resistant wave traps based on thermal region segmentation, characterized in that: Follow these steps: Step 1: Collect an infrared image A1 of the wave trap with a dimension of W×H, where W and H represent the width and height of the infrared image A1 respectively; Step 2, perform denoising on A1 to obtain a denoised infrared image A4; Step 2.1, after converting A1 into a grayscale image A2 of the same scale, A2 is processed using a column weighting and layered filtering algorithm to obtain an infrared image A3 without stripe noise; Step 2.2: After taking the logarithmic transformation of A3, use the Coiflet wavelet basis to perform N-layer wavelet decomposition on A3 to obtain the wavelet image set {G i |1≤i≤N}, where G i represents the wavelet image obtained by decomposition at the i-th layer; Step 2.3: Use the linear fitting algorithm based on SVD decomposition and information entropy to fit {G i |1≤i≤N} is processed to obtain an infrared image A4 without stripe noise and speckle noise; Step 3: Use the gradient weighted region growing method to segment the infrared image A4 to obtain a segmented infrared image A5; Step 4: According to the grayscale histogram of the segmented infrared image A5, a grayscale probability density function is obtained, and then the average grayscale value T of the wave trap area in A1 is obtained by using the grayscale probability density function. device , and then obtain the effective extreme point set of the wave trap area in the three-dimensional grayscale distribution diagram corresponding to A5; Step 5: Take each coordinate in the set of valid extreme points as the cluster center, and use the KNN algorithm to cluster the pixels in A5 to obtain the cluster information matrix. Then, according to the threshold condition, find the thermal anomaly area of ​​the wave trap from the cluster information matrix, and draw a mask at the corresponding position in A1 to obtain the thermal anomaly positioning image A6 of the wave trap.

2. The method for locating thermal faults of high noise resistant wave traps based on thermal region segmentation according to claim 1, characterized in that: The column weighting and layered filtering algorithm of step 2.1 comprises the following steps: Step 2.1.1, use the gradient operator to process the grayscale image A2 to obtain the horizontal gradient map Grad; Step 2.1.2, define the column weighted filter window: Define a filter window with a dimension of (2k+1)×(2k+1), where 2k+1 is the dimension of the filter window and k is a positive integer; The pixel value S(i, j) at the center position (i, j) of the filter window in A2 is obtained by using formula (1), and the boundary constraint condition of the filter window is set to: , , thereby sliding on the grayscale image A2 from left to right and from top to bottom according to the boundary constraint condition to obtain a sliding window image B1 containing low-frequency information; (1) In formula (1), represents the mean pixel value of the xth column in A2, and , represents the pixel value at coordinate (x, y) in A2; About A x The correction function of , represents an artificially determined regulatory factor, represents the mean pixel value of the x+1th column in A2, sgn(·) represents the sign function, represents the mean pixel value of the xth column in Grad, and , Represents the pixel value at coordinate (x, y) in Grad; represents the column weight of the x-th column in A2, and , σ is the variance of the set column weights; Step 2.1.3, after subtracting B1 from A2, a difference image B2 is obtained, and B2 is subjected to high-pass filtering in the column direction to obtain a high-frequency information image B3; Step 2.1.4: Use formula (2) to process B3 and obtain the corrected image B4: (2) In formula (2), B4(x, y) and B3(x, y) represent the pixel values ​​at coordinates (x, y) in B4 and B3 respectively; represents the average value of pixels in B3; λ is an artificially determined noise reduction factor; Step 2.1.5, add B1 and B4 to obtain the true grayscale image B5, and convert B5 to the same color space as A1 to obtain the infrared image A3 without stripe noise.

3. The method for locating thermal faults of high noise resistant wave traps based on thermal area segmentation according to claim 2 is characterized in that: The linear fitting algorithm based on SVD decomposition and information entropy in step 2.3 includes the following steps: Step 2.3.1: For the wavelet image set {G i | 1 ≤ i ≤ N}, each wavelet image is decomposed by SVD in the horizontal sub-band and vertical sub-band directions to obtain a set of non-zero singular value sequences {(Q i , P i )| 1 ≤ i ≤ N}, where Q i is the i-th wavelet image G i A sequence of non-zero singular values ​​in the horizontal subband, and Q i ={Q i,s | 1 ≤ s ≤S i }, Q i,s Indicates Q i The s-th horizontal non-zero singular value in S i Indicates Q i The total number of horizontal non-zero singular values ​​in P i is the i-th wavelet image G i The non-zero singular value sequence of the subband in the vertical direction, and P i ={P i,t | 1 ≤t ≤T i }, P i,t Indicates Q i The tth vertical non-zero singular value in , T i Indicates Q i The total number of vertical non-zero singular values ​​in ; Step 2.3.2, traverse {(Q i ,P i ) | 1 ≤ i≤ N}, remove Q in turn i , P i The non-zero singular values ​​less than the median are obtained, and the non-zero singular value sequence set after removal is obtained {(Q i ',P i ') | 1≤i≤N}; where Q i ' is the i-th wavelet image G i The non-zero singular value sequence after sub-band removal in the horizontal direction; P i ' is the i-th wavelet image G i The sequence of non-zero singular values ​​after subband removal in the vertical direction; Step 2.3.3, traverse {(Q i ',P i ') | 1≤i≤N}, calculate Q respectively i The average value of 'Q i,ave ' and P i The average value of ' i,ave ', and select Q i 'Middle is greater than Q i,ave The non-zero singular values ​​of ' and P i ' is greater than P i,ave After reconstructing the non-zero singular values ​​of ', we get the reconstructed wavelet image set {G i ' | 1 ≤ i ≤ N}; where G i ' represents the wavelet image after reconstruction of the i-th layer; Step 2.3.4: Calculate G i The information entropy H of the diagonal subband wavelet coefficients i,1 , and then use formula (3) to get the transformed information entropy H i,2 : (3) Step 2.3.5: Based on prior information, i,2 Perform polynomial fitting and calculate G i The noise variance of the diagonal subband is i : Step 2.3.6: Using the empirical Bayes formula and σ i Estimation {G i ' | 1 ≤ i ≤ N}, thus obtaining the real wavelet image set {G i ''|1≤i≤N}, where G i '' represents the real wavelet image of the i-th layer; Step 2.3.7, for {G i Each layer of the real wavelet image within ''|1≤i≤N} is successively subjected to exponential transformation and wavelet inverse transformation, and then image reconstruction is performed to obtain an infrared image A4 without stripe noise and speckle noise.

4. The method for locating thermal faults of high noise resistant wave traps based on thermal region segmentation according to claim 3 is characterized in that: The gradient weighted region growing method in step 3 comprises the following steps: Step 3.1: After converting A4 into a grayscale image A4' of the same scale, perform horizontal and vertical gradient detection on the grayscale image A4', and obtain the horizontal gradient image Grad x , longitudinal gradient map Grad y ; Step 3.2: After converting A4 to HSV space, use the watershed algorithm to pre-segment the image after HSV space conversion to obtain the first sub-region set ,in, represents the u-th first sub-region, and The set of adjacent sub-regions of , express The vth adjacent sub-region of express The total number of adjacent sub-regions of; Step 3.3, initialize u=1; Step 3.4, calculation and The root square difference of the saturation component between u , hue component root square difference simH u , and use formula (4) to calculate and The gradient cosquare root difference simG u ; (4) In formula (6), and Respectively The mean transverse gradient and the mean longitudinal gradient of and Represents sub-areas and the mean lateral gradient and the mean longitudinal gradient of all its adjacent sub-regions; Step 3.5: Calculate using formula (5) The similarity function f u If f u Less than threshold , then Mark it as the seed region and go to step 3.6; otherwise, go directly to step 3.6; (5) In formula (5), , , are the three weight coefficients respectively; Step 3.6: If u is less than sub1, assign u+1 to u and return to step 3.4 to execute sequentially; otherwise, the seed region set is obtained; Step 3.7, according to the seed growth rule, select all seed regions from the seed region set for growth to obtain a second sub-region set; Step 3.8: Perform rectangle fitting on all sub-regions in the second sub-region set, and remove the rectangles whose aspect ratio after fitting is greater than the threshold. sub-regions, and arrange the remaining sub-regions in descending order according to the distance from the geometric center to the center of A4 to obtain a third sub-region set. At the same time, arrange the remaining sub-regions in descending order according to their own area to obtain a fourth sub-region set; select the sub-region with the smallest sum of indexes in the third sub-region set and the fourth sub-region set as the segmented infrared image A5.

5. The method for locating thermal faults of high noise resistant wave traps based on thermal area segmentation according to claim 4 is characterized in that: The step 4 comprises the following steps: Step 4.1: Grayscale A5 and draw a grayscale histogram according to the grayscale value of the pixel points, so as to obtain the grayscale probability density function F(t) of A5 using the kernel estimation method, where t represents the grayscale value and t∈(T min ,T max ), where T min 、T max are the minimum and maximum values ​​in the grayscale histogram respectively; Step 4.2: Use the Otsu algorithm to obtain the segmentation threshold T of F(t), and take the pixel points in A5 with pixel values ​​less than T as background points, and the pixel points with pixel values ​​greater than T as the wave blocker area points; thus, use equations (6) and (7) to obtain the average grayscale value T of the wave blocker area respectively. device and the average gray value T of the background area env ; (6) (7) Step 4.3, draw a three-dimensional grayscale distribution diagram of A2 on a rectangular coordinate system, wherein the x-axis and y-axis of the rectangular coordinate system represent the horizontal coordinate and vertical coordinate of the image pixel point, and the z-axis represents the grayscale value corresponding to the pixel point; Define plane z0 and initialize it to z0=T device ; Step 4.4: Assign a value to z0. If there is a point tangent to z=z0 in the three-dimensional grayscale distribution diagram, store the x and y axis coordinates of the tangent point into the extreme point set and proceed to step 4.

5. Otherwise, proceed directly to step 4.

5. is the set parameter; Step 4.5: Determine whether z0 is less than T max If yes, go to step 4.4; otherwise, output the coordinate set of the effective extreme point {(x m ,y m )|1≤m≤E max }, where (x m ,y m ) represents the coordinates of the mth effective extreme point, E max Represents the total number of extreme points.

6. The method for locating thermal faults of high noise resistant wave traps based on thermal region segmentation according to claim 5, characterized in that: The step 5 comprises the following steps: Step 5.1: Take each coordinate in the effective extreme point set as the cluster center, use the KNN algorithm to cluster the points in the wave trap area, and obtain the cluster information matrix , where A max represents the total number of cluster regions, S e represents the area of ​​the e-th cluster region, T e Represents the average gray value of the e-th cluster area; represents the relative temperature difference of the e-th cluster area, and ; W e represents the weight of the e-th cluster region, and ; Step 5.2: Calculate the threshold T using formula (8) th : (8) In formula (8), is the correction factor; Step 5.3, initialize e=1; Step 5.4: If Greater than the threshold T th , then according to the position of the e-th cluster area in the infrared image A1, a mask is drawn at the corresponding position in the infrared image A1, and the process goes to step 5.5; otherwise, the process goes directly to step 5.5; Step 5.5: If e is less than A max , then e+1 is assigned to e, and the process goes to step 5.4 for sequential execution; otherwise, it means that an infrared image with a mask is obtained and output as the thermal anomaly positioning image A6.

7. An electronic device, comprising a memory and a processor, characterized in that: The memory is used to store a program that supports the processor to execute the high-noise-resistant wave trap thermal fault location method according to any one of claims 1 to 6, and the processor is configured to execute the program stored in the memory.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method for locating thermal faults of a high noise resistant wave trap according to any one of claims 1 to 6 are executed.

Images