Transformer iron core detection method based on computer vision

Abstract

The invention relates to the technical field of industrial component detection, in particular to a transformer iron core detection method based on computer vision, which comprises the following steps of: acquiring an iron core image, extracting key pixel characteristics, screening a directional scattering abnormal region to generate an interference map, extracting a consistent gradient region correction image to generate a reconstruction map, and positioning a symmetric disturbance generation structure map by integral gray difference. And analyzing an overlapping relation by a superposition structure graph to generate an abnormal component graph, and evaluating a risk level by matching a reference index to generate an early warning graph layer. Interference reflection and structural features can be distinguished through linkage analysis of the pixel direction vector and the brightness change frequency, correction of a distorted area in an image is realized based on a gray statistical stable value, and the distortion of the image is corrected by constructing a symmetric point map and analyzing the change trend of a gradient difference value sequence. And the structural overlapping relation is quantitatively judged by combining a component mapping profile diagram, so that the relevance between an abnormal region and a key component is clearly expressed, and the grading evaluation capability of various fault risks in the iron core is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of industrial component detection, and in particular to a transformer core detection method based on computer vision. BACKGROUND

[0002] Industrial component detection technology covers non-destructive testing methods based on electromagnetic, ultrasonic, eddy current, infrared, optical, laser, etc. principles, and also includes a technical system that uses sensor arrays, signal acquisition and processing systems, and intelligent algorithm models to continuously monitor and analyze state parameters of components during operation. Component detection is widely used in high-voltage equipment such as transformers, circuit breakers, cables, transformers, and insulators to improve the safety, maintainability, and accuracy of life prediction of equipment operation, and to avoid sudden failures and unplanned downtime.

[0003] The transformer core is one of the key components in power equipment, and its detection through industrial component detection technology can effectively avoid structural and technical defects and ensure its safety in use. Specifically, transformer core detection refers to a technical method for identifying and evaluating the structural state, electromagnetic performance, and defect conditions of the internal core components of a transformer. Its main uses include detecting whether there are problems such as looseness, misplacement, local overheating, abnormal clamping, and abnormal electromagnetic vibration in the core to prevent equipment risks such as increased eddy current loss, noise increase, and insulation breakdown caused by core failure. Through the detection method, the running state of the core can be accurately diagnosed to assist maintenance personnel in developing maintenance plans and improve the stability and safety of transformer operation.

[0004] Traditional detection methods rely only on electromagnetic performance or structural state signals for core evaluation, and cannot stably identify abnormal areas of the transformer core in environments with optical interference, structural obstruction, or unstable vibration frequency, especially when the clamping device is loose or the core laminations have a small displacement. Due to the lack of image analysis and symmetry modeling means, it is difficult to accurately capture local disturbance signals, resulting in an increase in abnormal misjudgment rate. For example, under light source interference conditions, a reflective area is misjudged as a defect site, or in a structural overlapping area, the influence range of a key component is missed, causing distortion of the risk assessment results. SUMMARY

[0005] The purpose of the present application is to solve the shortcomings in the prior art and provide a transformer core detection method based on computer vision.

[0006] The technical solution adopted by the present application to solve the above technical problems is: The transformer core detection method based on computer vision comprises the following steps: S1: Obtain the image of the transformer core, call the brightness gradient direction vector of each pixel and the brightness gradient change amplitude in the surrounding neighborhood, combine the gray-scale fluctuation frequency in the time series, filter the pixel areas with directional scattering anomalies, and generate the core reflection interference area map. S2: Based on the iron core reflection interference region map, extract the pixel grayscale set of regions with consistent gradients in three adjacent directions, replace the original abnormal region pixel grayscale value with the stable grayscale value, correct the image distortion caused by reflection or interference, and generate the iron core edge structure reconstruction image. S3: Call the edge pixel information of the longitudinal or transverse extension area of ​​the iron core in the image of the reconstructed image of the iron core edge structure, extract the corresponding continuous change point sequence, and filter it in combination with the gradient mean square error threshold to locate the boundary of the abnormal region that cannot converge to the symmetric trend, and generate the iron core symmetry perturbation structure map. S4: Based on the core symmetry disturbance structure diagram, calculate the percentage of overlap area between the abnormal area and the key component and the intersection length of the structural boundary, identify the abnormal area covering the key part, and generate a distribution diagram of the core abnormality affecting the component.

[0007] In a further technical solution of the present invention, the iron core reflection interference region map includes the boundary of the local brightness change region, the high-frequency pixel distribution with consistent direction, the brightness fluctuation sequence between image frames, and the suspicious reflection area identification mask; the iron core edge structure reconstruction image includes the gray-level correction block of the interference region, the gradient direction matching region, the stable gray-level region of the non-interference frame, and the image complete edge reconstruction region; the iron core symmetry disturbance structure map includes the symmetry mapping path annotation layer, the gray-level difference map of symmetry points, the structural offset response mask, and the continuous jump contour of the abnormal region; the iron core abnormality affected component distribution map includes the boundary projection map of key components, the overlapping label of abnormal components, the spatial index of the influence range, and the overlap rate distribution layer.

[0008] In a further technical solution of the present invention, the step of obtaining the spectrum of the iron core reflection interference region specifically includes: S111: Acquire images of the transformer core, extract pixel sets of the contact boundary of the clamping component, the lamination area of ​​the core, and the edge of the magnetic flux channel, monitor image frame data of pixels in the target area, calculate the inter-frame difference of the brightness value sequence of each pixel, and perform frequency statistics on the number of grayscale fluctuations in the time series to obtain the pixel brightness time change feature information. S112: Based on the pixel brightness time change feature information, call the brightness gradient direction vector value of the pixel in the current frame, and construct the set of angle vectors between the gradient direction of the pixel and the surrounding neighboring pixels. Compare whether the angle difference in the angle set is within the brightness direction concentration determination threshold, filter out pixels that meet the brightness gradient direction consistency condition, and obtain pixel direction consistency area information. S113: call the frequency value in the pixel direction consistency region information and the pixel brightness time variation characteristic information, jointly judge the time fluctuation frequency and direction concentration of each pixel point in the region, when the fluctuation frequency is higher than the high-frequency reflection identification threshold and the direction consistency meets the condition, mark the region as a reflection abnormal region, and generate a core reflection interference region map.

[0009] In a further technical solution of the application, the process of comparing whether the angle difference in the angle difference set is within the brightness direction concentration determination threshold is specifically: Call the brightness gradient direction vector value and the brightness gradient direction vector value corresponding to the eight neighboring pixel points, calculate the direction angle value between each group of center pixel points and neighboring pixel points, and generate an angle difference set including eight angle differences; According to the dispersion degree of each direction angle value in the angle difference set, calculate the maximum difference value and the standard deviation value of the angle difference set; When any one of the maximum difference value and the standard deviation value of the angle difference set is lower than the preset brightness direction consistency judgment threshold, it is determined that the center pixel point meets the brightness gradient direction consistency condition; the set value of the brightness direction consistency judgment threshold is less than 15 degrees S211: According to the abnormal pixel region marked in the core reflection interference region map, extract the pixel points in the adjacent region outside the boundary, call the gradient direction vector value of the pixel points in the current frame image, and calculate the angle difference value between the gradient direction and the center pixel gradient direction, select three direction regions with angle difference value not exceeding the brightness direction consistency judgment threshold, extract the pixel gray value respectively, and generate a direction consistency gray set; S212: Based on the pixel gray value sequence extracted from the direction consistency gray set, search the gray evolution track in the continuous image frames at the same spatial position, and eliminate the frames with instantaneous brightness surge, select the pixel gray sequence in the stable frame range, perform arithmetic average on the sequence, obtain the center stable gray value, calculate the difference between the maximum value and the minimum value of the sequence, construct the gray stability fluctuation factor, and generate a stable gray replacement value set. S213: Call the center gray value in the stable gray replacement value set to replace the pixel gray information of each abnormal region in the core reflection interference region map, optimize the region boundary continuity and texture consistency, integrate the replacement region on the original image, and output the reconstructed image frame sequence to generate a core edge structure reconstruction image.

[0010] In a further technical solution of the application, the acquisition step of the core symmetry disturbance structure map is specifically: S311: Call the edge pixels of the area extending along the longitudinal and transverse direction of the core in the core edge structure reconstruction image, collect the two-dimensional coordinate index and corresponding gray gradient value of the edge pixel points, and establish an equidistant point index pair based on the image center axis to construct a symmetric structure point mapping graph, and obtain the symmetric structure index distribution information; S312: Based on the equidistant point pair in the symmetric structure index distribution information, extract the gray gradient value sequence, perform difference operation and integral normalization according to the index order, and obtain the path symmetric offset of each symmetric scanning path by operation, filter the paths with path symmetric offset greater than the symmetry offset threshold, mark as structure imbalance area, and extract the jump point on the path to obtain the symmetric gradient offset point set; S313: Call the coordinate index of the marked point in the symmetric gradient offset point set, and construct a jump point clustering block based on the transverse and longitudinal adjacent features, filter the fragments with a span less than the region continuity judgment threshold, extract the edge point sequence with continuous distribution and stable offset direction from the center axis, and establish a core symmetry disturbance structure graph.

[0011] In a further technical solution of the application, the step of obtaining the core abnormal influence component distribution graph is specifically: S411: According to the abnormal region position coordinates marked in the core symmetry disturbance structure graph, extract the boundary index range of each abnormal region in the two-dimensional coordinate system, and retrieve the structure contour boundary graph of the clamping unit arrangement area, the magnetic flux channel and the insulation lap boundary of the corresponding image frame in the core structure database, superimpose the boundary graph and the abnormal region boundary graph to generate an abnormal component overlap index set; S412: Call the pixel coordinates recorded in the abnormal component overlap index set, calculate the boundary intersection length value between the abnormal region and the component region respectively, and count the proportion of the intersection region corresponding pixel point set in the component region, calculate the pixel area percentage of each type of component covered according to the area accumulation ratio, and generate abnormal component coverage index information; S413: Based on the coverage ratio and boundary intersection length of the component in the abnormal component coverage index information, filter the component region with coverage ratio greater than the influence judgment threshold, record the associated component name and position coordinate index of the corresponding abnormal region, and mark the key part influence area with association, and generate a core abnormal influence component distribution graph.

[0012] In a further technical solution of the application, the method further comprises the following steps: S5: call the abnormal type of the identified area in the core abnormal influence component distribution map and the covered component type, combine the corresponding brightness reflectance value, symmetry offset amplitude and key structure overlap ratio, the brightness reflectance value is the ratio of the average brightness value of the abnormal area to the average stable gray value of the surrounding normal area, interval matching is carried out with the commonly used risk assessment reference index in the transformer core detection application, label assignment is carried out according to the risk level interval to which the combined parameters belong, and a core multi-class abnormal early warning level layer is obtained; The core multi-class abnormal early warning level layer specifically refers to a risk level color zoning map, an abnormal type level label set, a component corresponding risk label set and a warning level interval distribution map.

[0013] In a further technical solution of the present application, the obtaining step of the core multi-class abnormal early warning level layer is specifically: S511: call the identified abnormal area in the core abnormal influence component distribution map, extract the abnormal type and the covered component type in each area, and respectively associate the characteristic indexes such as the brightness reflectance value, the symmetry offset amplitude and the structure overlap ratio corresponding to the area, construct a multi-dimensional attribute vector set, and generate an abnormal component feature parameter set; S512: according to the abnormal component feature parameter set, based on brightness, structure and area proportion, calculate the risk measurement value of each abnormal area, interval matching is carried out with the risk level interval boundary value set, the corresponding risk label type is marked, and an abnormal component risk level label set is generated; S513: call each risk level identification and the corresponding abnormal area coordinate information in the abnormal component risk level label set, bind the risk level label with the abnormal area in position and draw a risk layer, which is layered and superimposed in the core structure reference map, and a core multi-class abnormal early warning level layer is established.

[0014] Compared with the prior art, the present application has the following advantages: In the present application, through the linkage analysis of the pixel direction vector and the brightness change frequency, the interference reflection and the structure feature can be distinguished, the recognition ability of the local disturbance area is improved, the distortion area in the image is corrected based on the gray statistical stable value, the actual boundary profile of the core is effectively restored, the symmetry disturbance boundary is accurately captured by constructing the symmetry point map and analyzing the gradient difference value sequence change trend, and the association between the abnormal area and the key component is clearly expressed by quantitatively judging the structure overlap relationship in combination with the component mapping contour map, the abnormal early warning level is accurately divided through the multi-dimensional index matching of brightness reflection, symmetry offset and structure overlap, and the grading evaluation ability of the multi-class fault risk of the core is improved. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a flowchart illustrating the process of obtaining the spectrum of iron core reflection interference regions according to the present invention; Figure 3 This is a flowchart illustrating the process of obtaining a reconstructed image of the iron core edge structure according to the present invention; Figure 4 This is a flowchart illustrating the process of obtaining the core symmetry disturbance structure diagram according to the present invention; Figure 5 This is a flowchart illustrating the distribution of components affected by core anomalies according to the present invention. Figure 6 This is a flowchart illustrating the process of obtaining multiple anomaly warning level layers for iron cores according to the present invention. Detailed Implementation

[0016] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0017] like Figure 1 As shown, the computer vision-based transformer core detection method of the present invention includes the following steps: S1: Obtain the image of the transformer core, extract the pixel set of the contact boundary of the clamping component, the core lamination area, and the edge of the magnetic flux channel, call the brightness gradient direction vector of each pixel and the brightness gradient change amplitude in the surrounding neighborhood, combine the gray-scale fluctuation frequency in the time series, determine whether the direction concentration and high-frequency brightness fluctuation characteristics are satisfied at the same time, filter the pixel area with directional scattering anomaly, and generate the core reflection interference area map. S2: Based on the pixel regions marked in the iron core reflection interference region map, extract the pixel grayscale set of regions with consistent gradients in three adjacent directions, average the grayscale intensity values ​​of the regions in the non-interference frame, calculate the upper and lower limits of grayscale intensity fluctuation, replace the original abnormal region pixel grayscale values ​​with stable grayscale values, correct the image distortion caused by reflection or interference, and generate the iron core edge structure reconstruction image. S3: Call the edge pixel information of the longitudinal or transverse extension area of ​​the iron core in the image to reconstruct the iron core edge structure, construct a symmetrical point map, integrate and accumulate the gray-level gradient difference between equidistant point pairs, extract the corresponding continuous change point sequence, and filter it in combination with the gradient mean square error threshold to locate the boundary of abnormal regions that cannot converge to the symmetrical trend, and generate the iron core symmetry perturbation structure map. S4: Based on the coordinates of the abnormal area in the core symmetry disturbance structure diagram, superimpose the structural mapping contour diagram of the matching core internal clamping unit arrangement area, magnetic channel, and insulation overlap boundary, calculate the percentage of overlap area between the abnormal area and the key component and the intersection length of the structural boundary, identify the abnormal area covering the key part, and generate a core abnormality affected component distribution diagram. S5: Call the abnormal type of the identified area in the core abnormal influence component distribution map and the covered component type, combine the corresponding brightness reflectance value, symmetry offset amplitude and key structure coincidence ratio, the brightness reflectance value is the ratio of the average brightness value of the abnormal area to the average stable gray value of the surrounding normal area, respectively Interval matching with the risk assessment reference index commonly used in transformer core detection application, label assignment according to the risk level interval of the combined parameters, and obtain the transformer core multi-class abnormal warning level layer.

[0018] The core reflection interference region map includes local brightness mutation region boundary, direction consistency high frequency pixel distribution, image frame brightness fluctuation sequence and suspicious reflection area identification mask, the core edge structure reconstruction image includes interference region gray correction block, gradient direction matching area, non-interference frame stable gray area and image complete edge reconstruction area, the core symmetry disturbance structure map includes symmetry mapping path annotation layer, symmetry point gray difference map, structure offset response mask and abnormal area continuous jump contour, the core abnormal influence component distribution map includes key component boundary projection map, abnormal component overlap label, influence range space index and coincidence rate distribution layer, and the core multi-class abnormal warning level layer specifically refers to risk level color partition map, abnormal type level label group, component corresponding risk label set and warning level interval distribution map.

[0019] As shown in Figure 2 , the acquisition steps of the core reflection interference region map are specifically: S111: Obtain the transformer core image, extract the pixel set of the clamping component contact boundary, the core lamination region and the magnetic flux channel edge, monitor the image frame data of the pixel points in the target area, calculate the frame difference value of the brightness value sequence of each pixel point, and count the frequency of the gray fluctuation times in the time sequence to obtain the pixel brightness time variation characteristic information; The obtained transformer core image is a grayscale image with a resolution of 1920×1080 pixels. First, based on the preset structural mask information, the pixel set of the contact boundary of the clamping component is extracted. This set consists of pixels located in the image coordinate ranges (210, 300) to (210, 780) and (1710, 300) to (1710, 780). At the same time, the core lamination area is extracted, with pixel coordinates ranging from (215, 50) to (1705, 1030). Finally, the core lamination area is extracted... The magnetic flux channel edge is located within a rectangular region with pixel coordinates between (880, 50) and (1040, 1030). Within these three target regions, for each pixel, its luminance value data is continuously monitored across a 100-frame image sequence. Taking the pixel at coordinates (450, 600) as an example, its luminance value sequence over 10 consecutive frames is recorded as {128, 127, 129, 185, 188, 132, 130, 129, 192, 135}. This luminance value sequence is then processed... Frame-by-frame interpolation is performed, resulting in a sequence of nine interpolation values: {-1, 2, 56, 3, -56, -2, -1, 63, -57}. Next, the absolute value of each value in this interpolation sequence is statistically analyzed, and a grayscale fluctuation baseline is established. This baseline is based on the brightness value sequence of 50 normal, non-reflective pixels within 1000 frames under stable lighting conditions. The standard deviation of all sequences is calculated, and three times the average value is taken as the baseline. In this example, the brightness standard deviation of 50 pixels... The average standard deviation is 1.5, so the fluctuation baseline is set to 4.5. The absolute values ​​of the aforementioned difference sequence {1, 2, 56, 3, 56, 2, 1, 63, 57} are compared with 4.5, and the number of times the value exceeds the baseline is counted. In this example, there are 4 values ​​(56, 56, 63, 57) greater than 4.5. Therefore, the grayscale fluctuation frequency of this pixel is recorded as 4. This calculation is repeated for all pixels in the target area to finally obtain the brightness time change feature information of all monitored pixels.

[0020] S112: Based on the pixel brightness time change feature information, call the brightness gradient direction vector value of the pixel in the current frame, and construct the set of angle vectors between the gradient direction of the pixel and the surrounding neighboring pixels. Compare whether the angle difference in the angle set is within the brightness direction concentration determination threshold, filter out the pixels that meet the brightness gradient direction consistency condition, and obtain the pixel direction consistency area information. The process of comparing whether the angle differences in the set of included angles are within the threshold for determining the concentration of light direction is as follows: Call the brightness gradient direction vector value and the brightness gradient direction vector value corresponding to the eight neighboring pixels, calculate the direction angle value between each group of center pixels and neighboring pixels, and generate an angle difference set including eight angle differences. According to the discrete degree of each direction included angle value in the angle difference set, the maximum difference value and the standard deviation value of the angle difference set are calculated; When any one of the maximum difference value and the standard deviation value of the angle difference set is lower than a preset brightness direction consistency judgment threshold value, it is determined that the center pixel point satisfies the brightness gradient direction consistency condition; the set value of the brightness direction consistency judgment threshold value is less than 15 degrees; According to the acquired pixel brightness time variation characteristic information, pixel points with fluctuation frequency greater than 0 are screened out as objects to be analyzed, and the brightness gradient direction vector value of any one qualified central pixel point, for example, the pixel point with coordinates (450, 600), in the current image frame is called, and the Sobel gradients Gx and Gy of the pixel point in the x and y directions are calculated, wherein the Gx value is 48, the Gy value is -64, the gradient direction angle is calculated as atan2(-64, 48) and is about -53.1 degrees, and then a set of angle difference vectors between the gradient directions of the central pixel point and eight neighborhood pixel points around the central pixel point is constructed. The specific execution process is as follows: the brightness gradient direction vector values of the eight neighborhood pixel points (coordinates (449, 599) to (451, 601), excluding the central point) are calculated respectively, for example, the Gx value of the right neighborhood pixel point (451, 600) is 52, the Gy value is -60, the gradient direction angle is about -49.1 degrees, and the directional angle value between the neighborhood pixel point and the central pixel point is |-53.1-(-49.1)|, that is, 4.0 degrees. After the eight neighborhood pixel points are calculated in turn, an angle difference set including eight angle difference values is generated, for example, {5.2, 4.5, 4.0, 6.1, 5.5, 4.8, 6.5, 5.8} in degrees, and then according to the dispersion degree of each directional angle value in the angle difference set, the maximum difference value and the standard deviation value of the set are calculated. The maximum difference value is the maximum value in the set minus the minimum value, that is, 6.5-4.0=2.5 degrees, and the standard deviation value is obtained by the standard deviation calculation formula and is about 0.85 degrees. At this time, the calculated maximum difference value 2.5 degrees and the standard deviation value 0.85 degrees are compared with a preset brightness direction consistency judgment threshold value, and the set value of the brightness direction consistency judgment threshold value is 12 degrees. The setting process of the value is as follows: 1000 image samples of the iron core lamination region under the normal operation state of the transformer without reflection interference are collected, the same angle difference calculation is performed on the pixel points and their neighborhoods in each sample to form a database including a large number of maximum difference values and standard deviation values, the cumulative distribution function curves thereof are drawn respectively, and the value corresponding to the 90% cumulative probability is selected as the threshold value to ensure that most of the small directional changes caused by device normal vibration or sensor noise can be filtered out. Experimental data show that the maximum difference value of 90% of the samples is less than 13.5 degrees, and the standard deviation value is less than 12.8 degrees. In order to improve the strictness of screening, the smaller integer value 12 degrees between the two is taken as the final threshold value. The calculated maximum difference value 2.5 degrees and the standard deviation value 0.85 degrees are both less than 12 degrees, so it is determined that the central pixel point (450, 600) meets the brightness gradient direction consistency condition. After traversing all the high-frequency fluctuation pixel points, the pixel direction consistency region information is obtained.

[0021] S113: Call the frequency value in the pixel direction consistency region information and the pixel brightness time variation characteristic information, and jointly judge the time fluctuation frequency and direction concentration of each pixel point in the region. When the fluctuation frequency is higher than the high-frequency reflection identification threshold and the direction consistency meets the condition, mark the region as a reflection abnormal region, and generate a core reflection interference region map; Call the frequency value in the pixel direction consistency region information and the pixel brightness time variation characteristic information, and jointly judge the time fluctuation frequency and direction concentration of each pixel point in the region. When the fluctuation frequency is higher than the high-frequency reflection identification threshold and the direction consistency meets the condition, mark the region as a reflection abnormal region, and generate a core reflection interference region map;

[0022] As shown in Figure 3 , the acquisition step of the core edge structure reconstruction image is specifically: S211: According to the abnormal pixel region marked in the core reflection interference region map, extract the pixel points in the adjacent region outside the boundary, call the gradient direction vector value of the pixel points in the current frame image, and calculate the angle difference value between the gradient direction and the center pixel point. Filter three direction regions with an angle difference value not exceeding the brightness direction consistency judgment threshold, respectively extract the pixel gray value, and generate a direction consistency gray set; According to an abnormal pixel region marked in the core reflection interference region map, the range of the region is image coordinates (450, 600) to (460, 610), the pixel points in the adjacent region outside the boundary of the region are extracted, and the specific operation is that for each pixel point on the boundary of the abnormal region, a range of 5 pixels extending outward in the normal direction of the pixel point is selected as the adjacent region, for example, for the pixel point (455, 600) in the abnormal region, the adjacent region pixels are (455, 595) to (455, 599), the gradient direction vector values of these adjacent pixel points in the current frame image are called, and the included angle difference between each adjacent pixel point and the gradient direction of the center pixel point (455, 600) is calculated, wherein the gradient direction of the center pixel point is-52.5 degrees, the gradient direction of the adjacent pixel point (455, 598) is-54.2 degrees, and the included angle difference is 1.7 degrees. The included angle difference is compared with the brightness direction consistency judgment threshold value, the adjacent pixels with an included angle difference of not more than 12 degrees are screened out, and according to the positions of these pixels, the three direction regions with the largest number of pixels are selected, for example, after screening, it is found that the qualified adjacent pixels in the upper, upper left and left directions are the most, then the gray values of all qualified pixels in the three direction regions are extracted respectively, and finally the gray values are integrated to generate a direction consistency gray set.

[0023] S212: Based on the pixel gray value sequence extracted in the direction consistency gray set, the gray evolution trajectory of the same spatial position in the continuous image frames is retrieved, the frames with instantaneous brightness sudden increase are eliminated, the pixel gray sequence in the stable frame range is screened, the sequence is arithmetically averaged, the center stable gray value is obtained, the difference between the maximum value and the minimum value of the sequence is calculated, the gray stability fluctuation factor is constructed, and the stable gray replacement value set is generated; Based on the direction consistency gray set, which contains the gray values of a total of 75 pixel points from three direction areas, for these 75 pixel points, retrieve their respective gray evolution trajectories in the continuous 20 frames of images. Take the pixel point located at (455, 598) in the set as an example, its continuous 20 frames of gray value sequence is {135, 136, 134, 135, 210, 137, 136, …, 138}. In order to eliminate frames with instantaneous brightness surge, calculate the inter-frame difference value of the sequence, obtaining the difference value sequence {1, -2, 1, 75, -73, …}. At the same time, calculate the standard deviation of the entire sequence, which is 16.5. Set the surge criterion as the absolute value of the inter-frame difference value exceeding 3 times the standard deviation, i.e. 3*16.5=49.5. The values of 75 and -73 in the sequence correspond to the 5th and 6th frames, which are identified as having instantaneous brightness surge, and are therefore eliminated. Within the remaining 18 stable frames after elimination, filter out the gray sequence of the pixel point, for example {135, 136, 134, 135, 137, 136, …, 138}. Calculate the arithmetic mean of this stable sequence to obtain the central stable gray value as 136.2, rounded to 136. At the same time, calculate the difference between the maximum value 138 and the minimum value 134 in the stable sequence, obtaining a difference value of 4, which is the gray stability fluctuation factor. Repeat the above process for each pixel point in the direction consistency gray set, and finally generate a stable gray replacement value set containing 75 central stable gray values and 75 fluctuation factors.

[0024] S213: Call the central gray value in the stable gray replacement value set to replace the pixel point gray information of each abnormal region in the core reflection interference region map. Optimize the region boundary continuity and texture consistency, integrate the replacement region on the basis of the original image, and output the reconstructed image frame sequence to generate the core edge structure reconstruction image. The central stable gray value in the stable gray replacement value set is called, for example, for the 75 central stable gray values calculated with the pixel point (455, 600) as the center, the 75 values are weighted and averaged, and the weight is allocated according to the calculated gray stability fluctuation factor, the smaller the fluctuation factor, the more stable the historical gray value of the pixel point, and the higher the weight, and the weight calculation formula is W_i=(1 / F_i) / sum(1 / F_j), wherein F_i is the fluctuation factor of the i-th pixel, and it is set that the finally obtained weighted average gray value is 137, then the value is used to replace the gray information of the abnormal pixel point (455, 600) in the core reflection interference region map, and the weighted average gray value of the surrounding stable pixels of each pixel point in the abnormal region is used to replace the gray information of the abnormal pixel point in the abnormal region, and after the replacement of the gray information of all the pixels in the abnormal region is completed, the boundary of the replacement region is subjected to 8-neighbor mean filtering processing, the gray transition of the boundary pixels is smoothed, the continuity of the region boundary and the texture consistency are optimized, finally all the regions subjected to the gray replacement and boundary optimization are spliced and integrated back to the basis of the original image frame, and a reconstructed image frame is output, the operation is performed on the entire image sequence frame by frame, and finally a core edge structure reconstructed image sequence is generated.

[0025] As shown in Figure 4 , the acquisition step of the core symmetry disturbance structure is specifically: S311: Call the edge pixels of the region of the core edge structure reconstructed image along the longitudinal and transverse directions, collect the two-dimensional coordinate indexes and corresponding gray gradient values of the edge pixels, and establish an equidistant point index pair based on the image center axis to construct a symmetric structure point mapping diagram and obtain symmetric structure index distribution information; The image is reconstructed using the iron core edge structure. Edge pixels extending longitudinally and laterally from this image are extracted using the Canny edge detection operator. A Gaussian filter kernel size of 5x5, a low threshold of 50, and a high threshold of 150 are set to obtain accurate single-pixel width edges. The two-dimensional coordinate indices and corresponding grayscale gradient values ​​of all edge pixels are collected. The grayscale gradient values ​​are calculated using the Sobel operator in the x and y directions, and their modulus is taken. Subsequently, equidistant point index pairs are established based on the geometric center of the image (for a 1920x1080 image, the central axis is x=960). Taking the 400th row of the image (y=400) as an example, if an edge point coordinate is detected on the left edge of this row... If the coordinates are (220, 400), then the coordinates of its symmetrical point on the central axis x=960 should be (960+(960-220), 400), i.e. (1700, 400). The program will search for whether there is an edge point within a specified tolerance range (e.g., ±3 pixels) around (1700, 400). If it exists, for example, an edge point is detected at (1702, 400), then the two pixels (220, 400) and (1702, 400) will be established as a pair of equidistant point index pairs. This operation is performed on all horizontal and vertical edges to construct a complete symmetrical structure point mapping map. This map records the coordinate information of all successfully matched symmetrical point pairs, and finally obtains the symmetrical structure index distribution information.

[0026] S312: Based on the equidistant point pairs in the symmetric structure index distribution information, extract the gray-level gradient value sequence, perform difference calculations and integral normalization according to the index order, and adopt an improved formula: ; The path symmetry offset of each symmetric scan path is obtained through calculation. Paths with path symmetry offset greater than the symmetry offset threshold are filtered out and marked as structural imbalance areas. Jump points on the path are extracted to obtain the set of symmetric gradient offset points. in, Indicates the first The first path The grayscale gradient values ​​of each pixel at the symmetrical points on the left are obtained by calculating the Sobel gradient values ​​of the equidistant symmetrical points on the left side of the image frame. Indicates the first The first path The grayscale gradient value of each pixel at the symmetrical point on the right. Indicates the first The number of pixel pairs whose absolute value of the gray-level gradient difference in a path exceeds the threshold for determining a jump is defined as follows: the threshold is 1.5 times the average pixel gradient within the local window. This represents the number of pixel pairs on each symmetrical path. The symmetry offset of the path is represented. The setting method of the symmetry offset threshold is as follows: the corresponding value of each normal region sample path is extracted, a sample distribution curve is formed, the 95% position of the cumulative distribution function is taken as the threshold, and the normal fluctuation is excluded and the abnormal structure offset path is screened. Based on the obtained symmetry structure index distribution information, one transverse symmetry scanning path is selected, for example, all matched equidistant point pairs on the y=400 row in the image, the gray gradient value sequence of the point pairs is extracted to form the left gradient sequence And the right gradient sequence , wherein represents the path index, is the serial number of the pixel pair on the path, and there are pixel pairs on the path, and the difference operation is performed from 1 to 1480 in the index order and the integral normalization is performed, and the improved formula is used: ; The calculation is performed. The parameter in the formula represents the gray gradient value of the first pixel in the first path at the left symmetric point, which is obtained by calculating the Sobel gradient value of the left equidistant symmetric point in the image frame, and represents the gray gradient value of the corresponding point on the right side, is the number of pixel pairs in the first path whose gray gradient difference absolute value exceeds the jump determination reference, and the reference is set to 1.5 times of the average value of all pixel gradients in a 5x5 local window on the path, is the total number of pixel pairs on each symmetric path, and the final calculation result is the path symmetry offset. The calculation logic of the formula is that the numerator quantifies the overall asymmetry degree of the whole path by accumulating the absolute difference of the gradient values of the symmetric point pairs, the denominator is a normalization term, which uses the square root of the sum of squares of all gradient values to represent the overall gradient energy on the path, so as to eliminate the influence of the difference in edge sharpness of different regions on the result, and the multiplication term is a penalty term, which is increased when there is a "jump point" with a sharp gradient jump on the path, and the logarithmic term nonlinearly amplifies the final offset . In a specific example, for the path , , it is set that is obtained by calculation, and , , for The average gradient in the 5x5 window on the path is 25, so the jump determination reference is The number of pixel pairs on the path that satisfy is counted, and 10 pairs are found, so These values are substituted into the formula: ; The path symmetry offset is calculated After that, it is compared with the symmetry offset threshold value, which is set in the following way: 5000 symmetric scanning paths are randomly selected from 100 core image samples confirmed to have no structural defects, and the values of each of them are calculated to form a sample distribution curve. These values are sorted from small to large, and the value at the position after sorting is taken as the threshold value. The experimental statistics show that the value is 2.5, and in this example, The value is greater than the threshold value 2.5, so the path is marked as a structural imbalance area, and the coordinates of the 10 jump points (i.e., pixel pairs with gradient difference absolute value exceeding 37.5) calculated on the path are extracted to obtain the symmetric gradient offset point set. The result shows that path 400 has significant structural asymmetry. The usefulness of the formula lies in the introduction of the gradient energy term for normalization, which makes the symmetry evaluation not affected by the strength of the edge itself, and through the design of the logarithmic penalty term , the measurement result is more sensitive to local and severe structural mutations, which can effectively identify symmetry destruction caused by local deformation or foreign object obstruction.

[0027] S313: Call the coordinate index of the marked point in the symmetric gradient offset point set, and construct a jump point clustering block based on the horizontal and vertical adjacent features, filter fragments with a span less than the region continuity judgment threshold, extract edge point sequences that are continuously distributed and stable in the center axis offset direction, and establish a core symmetry disturbance structure diagram; The acquired symmetric gradient offset point set is called, which contains all the jump point coordinate indexes marked on the structural imbalance path, such as 10 jump points on path 400 and jump points on other imbalance paths. The lateral and longitudinal adjacent features based on these marked points are used to construct jump point clustering blocks. The specific implementation process is as follows: traverse each jump point, check whether there are other jump points in its 8-neighborhood range, if there are, classify these two points into a class, and continue to check the neighborhood of the newly added points until no new point is added to a certain class, thereby forming a clustering block. In this process, fragments with a span less than the region continuity judgment threshold value of 5 pixels need to be filtered out. The threshold value is set based on the minimum process size of the core lamination. Any isolated disturbance smaller than this size is not considered a structural defect. For example, a clustering block composed of 4 adjacent jump points is filtered out because its size is less than 5 pixels. The remaining clustering blocks are composed of 5 or more continuous jump points. Then, the edge point sequence in these clustering blocks that is continuously distributed and stable in the offset direction from the center axis is extracted. The offset direction stability judgment means that more than 90% of the points in a clustering block have consistent offset (i.e., the difference between the left gradient and the right gradient) directions (e.g., all have left gradient greater than right). Finally, these edge point sequences that meet the conditions are integrated to establish a core symmetry disturbance structure diagram.

[0028] As shown in Figure 5 , the core abnormality influence component distribution diagram acquisition step specifically includes: S411: According to the position coordinates of the abnormal regions marked in the core symmetry disturbance structure diagram, extract the boundary index range of each abnormal region in the two-dimensional coordinate system, and retrieve the structural contour boundary diagram of the clamping unit arrangement area, the magnetic flux channel, and the insulation lap boundary of the corresponding image frame in the core structure database. Superimpose the boundary diagram and the abnormal region boundary diagram to generate an abnormal component overlap index set. According to the abnormal region position coordinates marked in the core symmetry perturbation structure diagram, one of the abnormal regions is extracted, and the boundary index range thereof in the two-dimensional coordinate system is from 218 to 245 in the x direction and from 398 to 422 in the y direction. Then, the structure contour boundary diagram of the clamping unit arrangement area, the magnetic flux channel, and the insulating lap joint boundary corresponding to the current processing image frame is retrieved from the core structure database, and the database pre-stores the accurate position information of each component marked in the transformer design drawing, for example, the database shows that the range of the contour boundary of a clamping unit in the image is from 200 to 250 in the x direction and from 390 to 440 in the y direction. The boundary diagram (x: 218-245, y: 398-422) of the abnormal region is superimposed on the boundary diagram (x: 200-250, y: 390-440) of the clamping unit, and the pixel region overlapping in the coordinate system is found. The range of the overlapping region is from 218 to 245 in the x direction and from 398 to 422 in the y direction. All the pixel coordinates of the part of the overlapping region are recorded, and the abnormal component overlapping index set is generated.

[0029] S412: The pixel coordinates recorded in the abnormal component overlapping index set are called, the boundary intersection length value between the abnormal region and the component region is calculated respectively, and the proportion of the intersection region corresponding pixel point set in the component region is counted. According to the area accumulation proportion, the pixel area percentage covered by each type of component is calculated, and the abnormal component coverage index information is generated. The abnormal component overlapping index set is called, which records the pixel coordinates of the overlapping region between the abnormal region and the clamping unit region. The boundary intersection length value between the two regions is calculated respectively, which is realized by counting the number of continuous pixel points on the overlapping region boundary, for example, the intersection length is 108 pixels. The proportion of the intersection region (x: 218-245, y: 398-422) corresponding pixel point set in the component region (x: 200-250, y: 390-440) is counted. First, the pixel area of the intersection region is calculated, which is (245-218+1)*(422-398+1)=28*25=700 pixels. Then, the total pixel area of the clamping unit component is calculated, which is (250-200+1)*(440-390+1)=51*51=2601 pixels. According to the area accumulation proportion, the pixel area percentage covered by the abnormal region to the clamping unit is calculated, which is (700 / 2601)*100%=26.9%. The intersection length 108 pixels and the coverage 26.9% are recorded, and the abnormal component coverage index information of the abnormal region is generated.

[0030] S413: Based on the coverage ratio of the component in the abnormal component coverage index information and the boundary intersection length, the component area with a coverage ratio greater than the influence judgment threshold is screened out, the associated component name and position coordinate index of the corresponding abnormal area are recorded, and the key part influence area with correlation is marked, and an iron core abnormal influence component distribution map is generated; Based on the abnormal component coverage index information, the coverage ratio of the abnormal area to the clamping unit is 26.9%, and the boundary intersection length is 108 pixels. The component area with a coverage ratio greater than the influence judgment threshold is screened out. The influence judgment threshold is set according to the importance of different components. For key bearing and fixed components such as clamping units, the threshold is set to 15%. This setting refers to the requirements for structural integrity in the relevant equipment maintenance manual, that is, any unknown structure coverage exceeding 15% must be risk assessed. Since the calculated coverage ratio 26.9% is greater than the set influence judgment threshold 15%, the clamping unit area is screened out, the associated component name "left upper corner clamping unit" and position coordinate index of the associated abnormal area (x: 218-245, y: 398-422) are recorded, and the area is marked as a key part influence area with correlation. Finally, all the identified areas are integrated to generate an iron core abnormal influence component distribution map.

[0031] As shown in Figure 6 , the acquisition steps of the iron core multi-class abnormal early warning level layer are as follows: S511: Call the identified abnormal area in the iron core abnormal influence component distribution map, extract the abnormal type and covered component type in each area, and respectively associate the corresponding characteristic indexes such as brightness reflection ratio, symmetry offset amplitude, and structure coincidence ratio, the brightness reflection ratio is the ratio of the average brightness value of the abnormal area to the average stable gray value of the surrounding normal area, construct a multi-dimensional attribute vector set, and generate an abnormal component feature parameter set; Call the identified abnormal area in the iron core abnormal influence component distribution map, for example, the "left upper corner clamping unit" influence area, extract its abnormal type in the area, according to the judgment result, the type is "symmetry structure imbalance caused by surface reflection", and the covered component type is "clamping unit", and respectively associate the corresponding characteristic indexes, including: brightness reflection ratio, the value is calculated by the average brightness value of the abnormal area (for example, 195) and the average stable gray value of the surrounding normal area (for example, 135), the ratio is 195 / 135≈1.44; symmetry offset amplitude, the value is the calculated path symmetry offset The maximum value is 5.362; the structural overlap ratio is the calculated coverage of 26.9%. The heterogeneous index data are integrated to construct a multi-dimensional attribute vector in the form of {abnormal region ID: 001, abnormal type: symmetry imbalance, component type: clamping unit, brightness reflectance: 1.44, symmetry offset: 5.362, overlap ratio: 0.269}. This operation is performed on all identified abnormal regions to finally generate a set of abnormal component feature parameters.

[0032] S512: Based on the set of characteristic parameters of abnormal components, and taking into account brightness, structure, and region proportion, the following formula is used: ; The risk metric value of each abnormal region is obtained through calculation, and interval matching is performed with the risk level interval boundary value set. The corresponding risk label type is marked, and an abnormal component risk level label set is generated. in, Indicates the first Risk measurement value for each abnormal region The normalized value representing the brightness-reflectance ratio is obtained by linearly normalizing the ratio of the gray level at the center of the abnormal region to the average gray level of the corresponding background region. This represents the normalized value of the symmetry offset magnitude, obtained by using the path structure offset from the preceding steps. After normalization to the maximum value, This represents the normalized value indicating the overlap ratio of critical structures. It is obtained as the ratio of the area of ​​overlap between the abnormal region and the critical component structure to the total area of ​​the component. (Parameter) , These are adjustment factors for brightness and structural disturbance, used to adjust their contribution to the overall risk. The risk level interval boundary value set is obtained by extracting risk metrics from a large number of classified iron core anomaly samples. Establish its cumulative probability distribution curve, and set risk interval boundary points based on engineering experience, such as low risk, medium risk, and high risk, to form a segmented level standard value set; Based on the set of characteristic parameters of abnormal components, for the abnormal region with ID 001, the formula is used based on its brightness, structure, and region proportion: ; Perform risk metric calculation. (The formula contains...) Indicates the first Risk metric values ​​for each abnormal region is the normalized value of the brightness reflectance ratio, which is obtained by linearly normalizing the brightness reflectance ratio 1.44, and the setting of the normalization interval is based on the statistics of the brightness ratio in historical data, which is usually between 1.0 (no reflection) and 3.0 (strong specular reflection), and linearly mapping this interval to [0, 1], then , is the normalized value of the symmetry offset amplitude, which is obtained by maximum normalization of the path structure offset , and the maximum value calculated in all paths of the current whole image is set to 8.0, then , , is the normalized value of the key structure overlap ratio, i.e. the coverage 0.269, and and are the adjustment factors of brightness and structure disturbance, respectively, used to adjust their contribution to the overall risk, according to expert experience and historical failure data analysis, the harmfulness of structure deformation is usually higher than that of surface reflection, so , is set to make the weight of structure disturbance greater, the calculation logic of this formula is that the core term is the weighted sum of brightness risk and structure risk, representing the severity of the anomaly itself, and the square operation is performed on it to amplify the risk measurement value of high severity events, so that the discrimination degree is higher than that of low severity events, and the denominator then adjusts the risk value based on the overlap ratio of the abnormal area and the key component, the larger the overlap ratio , the larger the denominator, and the final risk value will be correspondingly reduced, this design reflects a risk adjustment mechanism, i.e. an anomaly with a wide impact range but low severity may have a lower risk than an anomaly with a small impact range but extremely high severity.

[0033] Substitute the above parameter values into the formula to calculate: ; The risk measurement value is calculated as 0.254, and then it is matched with the interval boundary value set of the risk level interval, which is obtained by extracting the risk measurement values of 500 core anomaly samples marked as "low risk", "medium risk", and "high risk" by experts in the database, drawing the cumulative probability distribution curve, and selecting 30% and 80% quantile points as the boundary points of the risk interval according to engineering experience, and the final boundary value set is: low risk ( ), medium risk ( ), and high risk ( ), since the calculated The value is between 0.2 and 0.6, so the abnormal area is marked as a “medium risk” type, and finally a set of abnormal component risk level labels is generated. The result shows that the abnormal area with ID 001 has a medium level of risk.

[0034] Table 1 shows a sample data summary for setting a risk level interval: Table 1 Risk level interval setting sample data summary table As shown in Table 1, by statistical analysis of a large number of classified samples, data support is provided for the division of risk levels, ensuring the objectivity and rationality of the boundary value setting.

[0035] S513: Call each risk level identifier in the abnormal component risk level label set and the corresponding abnormal area coordinate information, bind the risk level label with the abnormal area, and draw a risk layer, which is layered and superimposed on the core structure reference map to establish a core multi-class abnormal early warning level layer; Call the abnormal component risk level label set, which contains the risk level identifier (such as “medium risk”) of each abnormal area and the corresponding coordinate information (such as x: 218-245, y: 398-422), bind the risk level label with the abnormal area, and through a graphical drawing function, superimpose a semi-transparent yellow layer on the rectangular area with coordinates (218, 398) to (245, 422) to represent “medium risk” (where green represents low risk and red represents high risk). This process generates a risk layer, and finally, all similar generated risk layers are layered and superimposed on the pre-prepared core structure reference map, which is a clear, interference-free core standard image. Finally, a core multi-class abnormal early warning level layer is established and output.

[0036] The above is only a preferred embodiment of the present application, and does not limit the present application in other forms. Any simple modification and replacement of the above embodiment without departing from the technical solution content of the present application, according to the technical essence of the present application, still belongs to the protection scope of the technical solution of the present application.

Claims

1. A computer vision-based transformer core detection method, characterized in that, The method comprises the following steps: S1: obtaining a transformer core image, calling a brightness gradient direction vector of each pixel point and a brightness gradient change amplitude in a surrounding neighborhood, combining a gray scale fluctuation frequency in a time sequence, screening a pixel area with directional scattering anomalies, and generating a core reflection interference area atlas; S2: according to the core reflection interference area atlas, extracting a pixel gray scale set of three adjacent direction gradient consistent areas, replacing original abnormal area pixel gray scale values with stable gray scale values, correcting image distortion caused by reflection or interference, and generating a core edge structure reconstruction image; S3: calling edge pixel information of a core along a longitudinal or transverse extension area in the core edge structure reconstruction image, extracting a corresponding continuous change point sequence, and screening in combination with a gradient mean square deviation threshold, positioning an abnormal area boundary that cannot converge to a symmetric trend, and generating a core symmetry disturbance structure diagram; S4: according to the core symmetry disturbance structure diagram, calculating an overlapping area percentage between an abnormal area and a key component and a structure boundary intersection length, identifying an abnormal area covering a key part, and generating a core abnormal influence component distribution diagram.

2. The computer vision based transformer core detection method of claim 1, wherein, The core reflection interference area atlas comprises a local brightness mutation area boundary, a direction consistency high-frequency pixel distribution, an image frame-to-frame brightness fluctuation sequence, and a suspicious reflection area identification mask, the core edge structure reconstruction image comprises an interference area gray scale correction block, a gradient direction matching area, a non-interference frame stable gray scale area, and an image complete edge reconstruction area, the core symmetry disturbance structure diagram comprises a symmetric mapping path annotation layer, a symmetric point gray scale difference diagram, a structure offset response mask, and an abnormal area continuous jump contour, and the core abnormal influence component distribution diagram comprises a key component boundary projection diagram, an abnormal component overlapping label, an influence range space index, and an overlapping rate distribution layer.

3. The computer vision-based transformer core detection method of claim 2, wherein, The acquisition step of the core reflection interference area atlas is specifically as follows: S111: obtaining a transformer core image, extracting a pixel set of a clamping component contact boundary, a core lamination area, and a magnetic flux channel edge, monitoring image frame data of a pixel point in a target area, performing frame-to-frame difference calculation on a brightness value sequence of each pixel point, and performing frequency statistics on a gray scale fluctuation frequency in a time sequence to obtain pixel brightness time variation characteristic information; S112: according to the pixel brightness time variation characteristic information, calling a brightness gradient direction vector value of a current frame of a pixel point, and constructing an included angle vector set between the gradient direction of the pixel point and the gradient direction of a surrounding neighborhood pixel point, comparing whether an angle difference in the included angle set is within a brightness direction concentration degree determination threshold, screening a pixel point meeting a brightness gradient direction consistency condition, and obtaining pixel direction consistency area information; S113: calling the pixel direction consistency area information and a frequency value in the pixel brightness time variation characteristic information, jointly judging a time fluctuation frequency and a direction concentration degree of each pixel point in a region, when the fluctuation frequency is higher than a high-frequency reflection identification threshold and the direction consistency meets a condition, marking the region as a reflection abnormal area, and generating a core reflection interference area atlas.

4. The computer vision based transformer core detection method of claim 3, wherein, The process of comparing whether an angle difference in the included angle set is within a brightness direction concentration degree determination threshold is specifically as follows: The direction included angle value between each group of center pixel points and the neighborhood pixel points is calculated by calling the luminance gradient direction vector value of the eight neighborhood pixel points corresponding to the luminance gradient direction vector value of the calling luminance gradient direction vector value, and an angle difference set including eight angle differences is generated; The maximum difference value and the standard deviation value of the angle difference set are calculated according to the discrete degree of each direction included angle value in the angle difference set; When any one of the maximum difference value and the standard deviation value of the angle difference set is lower than a preset luminance direction consistency judgment threshold value, it is determined that the center pixel point satisfies the luminance gradient direction consistency condition; the set value of the luminance direction consistency judgment threshold value is less than 15 degrees.

5. The computer vision based transformer core detection method of claim 4, wherein, The acquisition step of the iron core edge structure reconstruction image is specifically: S211: According to the abnormal pixel region marked in the iron core reflection interference region map, the pixel points in the adjacent region outside the boundary are extracted, the gradient direction vector value of the pixel points in the current frame image is called, and the included angle difference value between the gradient direction of the center pixel point is calculated, three direction regions with an included angle difference value not exceeding the luminance direction consistency judgment threshold value are screened, the pixel gray value is extracted respectively, and a direction consistency gray set is generated; S212: Based on the pixel gray value sequence extracted from the direction consistency gray set, the gray evolution track of the same spatial position in the continuous image frame is searched, the frame with instantaneous luminance sudden increase is eliminated, the pixel gray sequence in the stable frame range is screened, the sequence is arithmetically averaged, the center stable gray value is obtained, the difference between the maximum value and the minimum value of the sequence is calculated, the gray stability fluctuation factor is constructed, and a stable gray replacement value set is generated; S213: The center gray value in the stable gray replacement value set is called to replace the pixel gray information of each abnormal region in the iron core reflection interference region map, the region boundary continuity and texture consistency are optimized, the replaced region is integrated on the basis of the original image, and the reconstructed image frame sequence is output, and an iron core edge structure reconstruction image is generated.

6. The computer vision-based transformer core detection method of claim 5, wherein, The acquisition step of the iron core symmetry disturbance structure diagram is specifically: S311: The region edge pixels of the iron core along the longitudinal and transverse directions in the iron core edge structure reconstruction image are called, the two-dimensional coordinate index and the corresponding gray gradient value of the edge pixel points are collected, and the equidistant point index pair is established based on the image center axis to construct a symmetric structure point mapping diagram, and the symmetric structure index distribution information is obtained; S312: Based on the equidistant point pairs in the symmetric structure index distribution information, the gray gradient value sequence is extracted, the difference value operation is performed in the index order and is integrated and normalized, the path symmetric offset of each symmetric scanning path is obtained by operation, the paths with a path symmetric offset greater than a symmetry offset threshold value are screened and marked as structure imbalance regions, and the jumping points on the paths are extracted to obtain a symmetric gradient offset point set; S313: The coordinate index of the marked point in the symmetric gradient offset point set is called, and the jumping point clustering block is constructed based on the transverse and longitudinal adjacency characteristics, the fragments with a span less than a region continuity judgment threshold value are filtered, the edge point sequence with continuous distribution and stable offset direction from the center axis is extracted, and an iron core symmetry disturbance structure diagram is established.

7. The computer vision-based transformer core detection method of claim 6, wherein, The core abnormality influence component distribution map acquisition step specifically comprises the following steps: S411: According to the abnormal region position coordinates marked in the core symmetry disturbance structure map, the boundary index range of each abnormal region in the two-dimensional coordinate system is extracted, and the structure contour boundary map of the clamping unit arrangement area, the magnetic flux channel and the insulation lap joint boundary of the corresponding image frame in the core structure database is retrieved, the boundary map and the abnormal region boundary map are superimposed, and the abnormal component overlapping index set is generated; S412: The pixel coordinates recorded in the abnormal component overlapping index set are called, the boundary intersection length value between the abnormal region and the component region is calculated respectively, the proportion of the intersection region corresponding pixel point set in the component region is counted, the pixel area percentage of each type of component covered is calculated according to the area accumulation proportion, and the abnormal component coverage index information is generated; S413: Based on the coverage proportion and the boundary intersection length of the component in the abnormal component coverage index information, the component region with a coverage proportion greater than the influence judgment threshold is screened, the associated component name and position coordinate index of the corresponding abnormal region are recorded, and the key part influence area with association is marked as a key part influence area, and a core abnormality influence component distribution map is generated.

8. The computer vision-based transformer core detection method of claim 7, wherein, The method further comprises the following steps: S5: The abnormal type and the covered component type of the identified area in the core abnormality influence component distribution map are called, the corresponding brightness reflectance value, the symmetry offset amplitude and the key structure coincidence proportion are combined, the brightness reflectance value is the ratio of the average brightness value of the abnormal region to the average stable gray value of the surrounding normal region, interval matching is performed with the commonly used risk assessment reference index in the transformer core detection application respectively, label assignment is performed according to the risk level interval to which the combined parameters belong, and a core multi-class abnormality early warning level layer is obtained; The core multi-class abnormality early warning level layer specifically refers to a risk level color zoning map, an abnormal type level label set, a component corresponding risk label set and a warning level interval distribution map.

9. The computer vision-based transformer core detection method of claim 8, wherein, The core multi-class abnormality early warning level layer acquisition step specifically comprises the following steps: S511: The abnormal region marked in the core abnormality influence component distribution map is called, the abnormal type and the covered component type in each region are extracted, and the feature indexes such as the brightness reflectance value, the symmetry offset amplitude and the structure coincidence proportion corresponding to the region are associated respectively, a multi-dimensional attribute vector set is constructed, and an abnormal component feature parameter set is generated; S512: According to the abnormal component feature parameter set, the risk measurement value of each abnormal region is calculated based on the brightness, structure and area proportion, interval matching is performed with the risk level interval boundary value set, the corresponding risk label type is marked, and an abnormal component risk level label set is generated; S513: Each risk level label in the abnormal component risk level label set and the corresponding abnormal region coordinate information are called, the risk level label and the abnormal region are positionally bound and a risk layer is drawn, the risk layer is layered and superimposed in the core structure reference map, and a core multi-class abnormality early warning level layer is established.

Images