Transformer fault diagnosis method and system based on image recognition

Abstract

The application relates to the technical field of power equipment state monitoring, and particularly discloses a transformer fault diagnosis method and system based on image recognition, which comprises the following steps: collecting a transformer multi-modal image sequence in real time, screening through clarity and component integrity evaluation to form an initial image set; performing multi-scale space registration on the initial image set and a transformer normal state standard template to generate a reference image, and inversely deducing a displacement vector field based on pixel-level differences; under the constraint of a transformer structure geometry, smoothing, optimizing and geometrically reconstructing the displacement vector field to generate a structure true correction image; strengthening a fault feature in the gradient domain of the correction image, identifying a fault area through multi-channel feature extraction and matching with a transformer typical fault feature library, and generating a diagnosis report of comprehensive fault types, confidence and geometric parameters; the application effectively solves the problem of image geometric deformation caused by differences in shooting conditions, and improves the accuracy and reliability of fault identification.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power equipment state monitoring, in particular to a transformer fault diagnosis method and system based on image recognition. BACKGROUND

[0002] In recent years, fault detection technology based on image recognition has been gradually applied to power equipment state monitoring. As a key device in the power system, the running state of the transformer directly affects the reliability and safety of the power grid. Traditional transformer fault diagnosis mainly relies on methods such as periodic manual inspection, oil chromatographic analysis, and partial discharge detection. Manual inspection has strong subjectivity, low efficiency, and cannot achieve continuous monitoring; oil chromatographic analysis can effectively detect internal latent faults, but the response is lagging and the fault location cannot be determined; partial discharge detection is not sensitive to external mechanical faults.

[0003] The existing image recognition method directly analyzes the original image, and its diagnosis accuracy is subject to perspective distortion and spatial position deviation in the image acquisition process, resulting in inconsistent shape and size of the same component in different images, which reduces the reliability of fault judgment based on size measurement and shape features. To solve this problem, the present application establishes a multi-scale spatial registration of the transformer normal state standard template and the collected image, reversely deduces the displacement vector field representing local deformation, and optimizes and reconstructs the displacement field under the physical constraints of the transformer structure, to generate a corrected image with a real structure, which fundamentally eliminates the geometric errors introduced by the acquisition angle and distance, and establishes a stable and reliable image basis for accurate extraction and recognition of subsequent fault features. SUMMARY

[0004] The purpose of the present application is to provide a transformer fault diagnosis method based on image recognition to solve the problems in the background.

[0005] The purpose of the present application can be achieved by the following technical solutions:

[0006] The transformer fault diagnosis method based on image recognition comprises the following steps:

[0007] S1: Real-time acquisition of multi-modal image sequences of transformer components, and automatic screening of effective images containing key fault features based on image clarity and component integrity evaluation to form an initial image set;

[0008] S2: Multi-scale spatial registration of the initial image set and the pre-constructed transformer normal state standard template to generate a structure-aligned reference image;

[0009] S3: Reverse deduction of a displacement vector field representing local deformation of the image based on the pixel-level difference between the initial image set and the reference image;

[0010] S4: Based on the displacement vector field, the structure of the transformer is reconstructed to generate a corrected image with a real structure;

[0010] S4: under the constraint of transformer structure geometry, smoothing optimization is performed on the displacement vector field, and based on the optimized displacement field, the initial image is geometrically reconstructed to generate a structure-realistic corrected image;

[0011] S5: fault feature enhancement is performed in the gradient domain of the corrected image, abnormal region features are extracted, the fault region is identified according to the abnormal region features, and a final transformer fault diagnosis report is generated.

[0012] As a further scheme of the application: the initial image set is specifically constructed, and includes:

[0013] A fixed visible light camera and an infrared thermal imager are deployed around the transformer to synchronously trigger the acquisition of a visible light image sequence and an infrared image sequence of the transformer components, thereby forming an original multi-modal image data stream;

[0014] For each frame of image in the original multi-modal image data stream, a variance calculation based on second-order differential features is used to evaluate the image sharpness, and by comparing with a preset dynamic threshold, the frames with qualified sharpness are screened out to generate a preliminary screening image set;

[0015] Based on the prior three-dimensional geometric information of the transformer components, a standard contour projection template of the key components is constructed, each image in the preliminary screening image set is spatially registered with the template, the overlapping ratio of the visible region of the components is calculated as the integrity score, and the images with a score lower than the integrity threshold are removed.

[0016] Multi-scale gradient domain analysis is performed on the images that pass the integrity evaluation, potential fault features are preliminarily identified by extracting high-frequency gradient features and statistically analyzing their distribution abnormalities, and finally these images are collected as effective images to constitute the initial image set.

[0017] As a further scheme of the application: the structure-aligned reference image is specifically generated, and includes:

[0018] The anatomical semantic key points of the transformer components of each image are extracted from the initial image set, including the sleeve flange edge joint points, the heat sink rib plate root connection points, and the oil tank weld feature intersection points.

[0019] A Gaussian difference pyramid representation of the transformer normal state standard template is constructed, and a scale-invariant spatial coordinate set of the anatomical semantic key points is established at each pyramid level.

[0020] The anatomical semantic key points are bidirectionally and consistently matched with the scale-invariant spatial coordinate set, and the optimal affine transformation parameters are solved by an iterative closest point framework.

[0021] The optimal affine transformation parameters are used to perform global geometric transformation on the initial image set, so that the transformer component structure in the image is accurately aligned with the standard template in space, and a reference image with structure alignment is output.

[0022] As a further scheme of the present application, the displacement vector field representing local deformation of the image is reversely derived, and specifically includes:

[0023] A dense pixel correspondence relationship is established between the reference image and the corresponding image of the initial image set, local similarity of the image is quantified by calculating a normalized cross-correlation value in a neighborhood of each pixel, and an initial similarity distribution map is generated;

[0024] Anisotropic diffusion filtering is performed on the initial similarity distribution map, connectivity of a high-similarity region is enhanced along an image gradient direction, and local extreme values caused by noise are suppressed, so as to form an optimized similarity distribution map;

[0025] Based on the optimized similarity distribution map, a spatial offset vector of each pixel to a corresponding maximum similarity point is calculated by tracking a maximum similarity path in the initial image, starting from each pixel position in the reference image;

[0026] The spatial offset vectors of all pixels are reorganized according to their original positions, and a complete displacement vector field is constructed.

[0027] As a further scheme of the present application, the structure-true corrected image is generated, and specifically includes:

[0028] Regions corresponding to rigid structure components of the transformer, including a bushing flange and an oil tank wall joint, in the displacement vector field are identified, and rigid motion constraints are applied to these regions to maintain their geometric invariability;

[0029] Structure continuity constraints are established for regions corresponding to non-rigid connectors in the displacement vector field, a smooth displacement field satisfying the physical structure characteristics of the transformer is obtained by solving a partial differential equation group satisfying the equilibrium conditions of elastic mechanics;

[0030] Based on the smooth displacement field, a dense coordinate mapping relationship from the initial image to the reference image is established, and the initial image is resampled, so that geometric distortion is eliminated while the integrity of texture features is maintained;

[0031] A corrected image conforming to anatomical reality is generated by verifying the consistency of geometric dimensions of key components of the transformer in the reconstructed image with the standard template.

[0032] As a further scheme of the present application, the smooth displacement field satisfying the physical structure characteristics of the transformer is obtained by solving a partial differential equation group satisfying the equilibrium conditions of elastic mechanics, and specifically includes:

[0033] Anisotropic elastic parameter fields of the non-rigid connection region are established based on material properties of each component of the transformer;

[0034] Balance equations based on stress-strain relationship are constructed in the non-rigid connection region of the displacement vector field, and continuous partial differential equations are discretized into linear equations by finite difference method;

[0035] The linear equations are solved, the low-frequency components of the error are quickly eliminated in the coarse grid layer, and the high-frequency components of the error are accurately corrected in the fine grid layer, so that the displacement field correction quantity satisfying the mechanical balance condition is obtained;

[0036] The displacement field correction quantity and the original displacement vector field are weighted and fused to generate a smooth displacement field that maintains geometric deformation characteristics and satisfies the mechanical continuity of the transformer structure.

[0037] As a further scheme of the present application, the balance equations based on stress-strain relationship are constructed in the non-rigid connection region of the displacement vector field, and the continuous partial differential equations are discretized into linear equations by finite difference method, which specifically includes:

[0038] Based on the anisotropic elastic parameter field, the constitutive relationship between the stress tensor and the strain tensor is established at each pixel point in the non-rigid connection region, wherein the strain tensor is determined by the first-order spatial differentiation of the displacement vector field;

[0039] The stress tensor of each pixel point is substituted into the balance equation to obtain a partial differential equation about the second-order spatial differentiation of the displacement vector field;

[0040] The second-order differentiation term in the partial differential equation is discretized to establish a linear dependence relationship between the displacement vector of each pixel point and the displacement vector of its adjacent pixel point;

[0041] All pixel points in the non-rigid connection region are traversed, and the discretized equation is assembled into a large sparse linear equation group containing all unknown displacement vectors according to the spatial position sequence.

[0042] As a further scheme of the present application, the abnormal region is identified according to the characteristics of the abnormal region, and a final transformer fault diagnosis report is generated, which specifically includes:

[0043] In the corrected image after gradient domain enhancement, multi-channel feature extraction is performed on each candidate abnormal region, and the multi-channel features include hot spot distribution topological structure, texture spectrum distribution entropy value and boundary tortuosity quantization parameters of the region;

[0044] The extracted multi-channel features are bidirectionally matched with the typical transformer fault feature library, the Mahalanobis distance in the feature space is calculated to determine the preliminary fault type result, and the confidence level is evaluated;

[0045] Based on the primary election results, the adaptive morphological closing operation is carried out on the candidate region, the broken boundary is fused and the internal hole is filled, and the complete fault region three-dimensional geometric representation is generated;

[0046] The final diagnosis report containing fault qualitative conclusion and quantitative analysis data is automatically generated according to the preset diagnosis report template by comprehensively considering the fault type, confidence level, geometric parameter and spatial position information.

[0047] As a further scheme of the application, the extracted multi-channel feature is bidirectionally matched with the transformer typical fault feature library, the Mahalanobis distance in the feature space is calculated to determine the fault type primary election result, and the confidence level is evaluated, and specifically comprising:

[0048] Based on the multi-channel feature, the feature vector of the candidate abnormal region is constructed, and the reference fault feature vector with the same feature dimension is searched in the transformer typical fault feature library;

[0049] The Mahalanobis distance of the candidate feature vector and each reference fault feature vector in the feature space is calculated, and the Mahalanobis distance is obtained according to the covariance relationship between different feature dimensions;

[0050] The first three reference fault types with the smallest Mahalanobis distance are selected as the primary election result, and the corresponding distance value is recorded as the matching degree index;

[0051] Based on the Mahalanobis distance distribution characteristics of the three primary election results, the confidence level of each primary election result is calculated through distance ratio analysis and relative position evaluation, and a fault type primary election result set with confidence weight is formed.

[0052] The transformer fault diagnosis system based on image recognition comprises:

[0053] A dynamic sequence acquisition and quality screening module is used for real-time acquisition of multi-modal image sequences of transformer components, and automatically screens out effective images containing key fault features based on image clarity and component integrity evaluation to form an initial image set;

[0054] A multi-scale template registration module is used for multi-scale space registration of the initial image set and the pre-constructed transformer normal state standard template to generate structure-aligned reference images;

[0055] A deformation field reverse derivation module is used for reverse derivation of displacement vector field representing local deformation of images based on pixel-level difference between the initial image set and the reference image;

[0056] A physical constraint geometric correction module is used for smoothing optimization of the displacement vector field under the geometric constraint of the transformer structure, and geometric reconstruction of the initial image based on the optimized displacement field to generate structure-realistic corrected images;

[0057] A fault diagnosis and report generation module is configured to enhance fault features in a gradient domain of the corrected image, extract abnormal region features, identify a fault region based on the abnormal region features, and generate a final transformer fault diagnosis report.

[0058] Advantages of the present application:

[0059] (1) By establishing a transformer normal state standard template and a multi-scale space registration method, combining displacement vector field derivation and geometric reconstruction under physical constraints, the image geometric distortion caused by shooting angle and distance change is effectively eliminated, making the transformer images collected at different times and under different conditions comparable, significantly improving the consistency of size measurement and morphological feature analysis, and providing a geometric standardized analysis basis for fault diagnosis.

[0060] (2) Multi-modal image sequence analysis and gradient domain fault feature enhancement technology is adopted, multi-channel feature extraction such as hot spot distribution topological structure and texture spectrum distribution entropy value is adopted, combined with feature matching and confidence evaluation mechanism based on Mahalanobis distance, sensitive detection and accurate identification of early fault features of the transformer are realized, which can effectively distinguish normal state from potential fault, and improve the accuracy and reliability of fault diagnosis. BRIEF DESCRIPTION OF DRAWINGS

[0061] The present application will be further described below with reference to the accompanying drawings.

[0062] Figure 1 is a flow chart of the method of the present application;

[0063] Figure 2 is a system block diagram in the present application. DETAILED DESCRIPTION

[0064] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0065] Please refer to Figure 1 The present application is a transformer fault diagnosis method based on image recognition, which includes the following steps:

[0066] S1: Real-time acquisition of multi-modal image sequence of transformer components, and automatic screening of effective images containing key fault features based on image clarity and component integrity evaluation to form an initial image set;

[0067] S2: Multi-scale spatial registration of the initial image set with the pre-constructed transformer normal state standard template to generate a structure-aligned reference image;

[0068] S3: Based on the pixel-level difference between the initial image set and the reference image, the displacement vector field representing the local deformation of the image is reversely deduced;

[0069] S4: Under the geometric constraint of the transformer structure, the displacement vector field is smoothed and optimized, and the initial image is geometrically reconstructed based on the optimized displacement field to generate a structure-realistic corrected image;

[0070] S5: In the gradient domain of the corrected image, the fault feature is strengthened, the abnormal area feature is extracted, the fault area is identified according to the abnormal area feature, and the final transformer fault diagnosis report is generated.

[0071] In S1, the multi-modal image sequence of the transformer component is collected in real time, and the effective images containing key fault features are automatically screened based on image clarity and component integrity evaluation to form an initial image set, which specifically includes:

[0072] Through the fixed visible light camera and infrared thermal imager deployed around the transformer, the visible light image sequence and infrared image sequence of the transformer component are collected in a synchronous triggering manner. The visible light camera collects the detailed information of the surface state of the transformer, and the infrared thermal imager synchronously acquires the temperature distribution information of the same part, forming a raw multi-modal image data stream aligned in time and space.

[0073] The clarity of each frame of image in the raw multi-modal image data stream is evaluated. A clarity evaluation method based on Laplacian operator is adopted to perform second-order differential processing on the image, and the variance value of the processed image is taken as the clarity evaluation value. A dynamic threshold adjustment mechanism is set, which adjusts the threshold value according to the environmental lighting conditions and shooting distance. The image frames with clarity evaluation values lower than the threshold value are removed, and the image frames meeting the clarity requirement are retained to generate a preliminary screening image set.

[0074] Based on the prior three-dimensional geometric information of the transformer component, a standard contour projection template of the key component is constructed. The template contains the standard shape contour of the main components such as transformer bushing, cooling fin, oil tank, etc. Each image in the preliminary screening image set is spatially registered with the standard contour projection template at the corresponding angle, and the overlap ratio of the visible area of the component in the actual shooting image with the template contour is calculated as the component integrity score. The integrity threshold is set to 85%, and the images with integrity scores lower than the threshold are removed.

[0075] Multi-scale gradient domain analysis is performed on the images that pass the integrity assessment. By constructing a Gaussian pyramid of the image, high-frequency gradient features of the image are extracted at 3 different scale spaces, the distribution characteristics of the gradient features at each scale are counted, and the skewness and kurtosis of the gradient distribution are calculated as abnormality assessment indicators. When the absolute value of the skewness of the gradient distribution is greater than 0.5 or the kurtosis exceeds the range of 3.2-4.8, it is determined that the image contains potential fault features. All images determined to contain potential fault features are collected to form an initial image set for subsequent analysis.

[0076] In S2, the initial image set is registered with the pre-constructed transformer normal state standard template in multi-scale space to generate a structure-aligned reference image, specifically including:

[0077] The transformer component dissection semantic key points of each image are extracted from the initial image set. These key points include the sleeve flange edge joint points, the fin rib root connection points, and the oil tank weld feature intersection points. The key point extraction process uses a gradient direction histogram-based feature detection method to identify the key point positions with stable geometric characteristics by calculating the gradient direction distribution statistical features in the local region of the image. For each candidate key point, the response value in the scale space is calculated, and the stable key points with response values exceeding the preset threshold are retained to finally form the dissection semantic key point set of each image.

[0078] A Gaussian difference pyramid representation of the transformer normal state standard template is constructed. A pyramid structure containing 5 scale levels is established, and the images at each scale are convolved by a Gaussian kernel function with a scale factor increasing by a factor of 1.2. At each level of the Gaussian difference pyramid, according to the geometric characteristics of the template image, a set of scale-invariant spatial coordinates of the dissection semantic key points is established. These coordinate sets are normalized by the relative position relationship of the key points at different scales to ensure their scale invariance.

[0079] The dissection semantic key points extracted from the initial image are matched with the scale-invariant spatial coordinate set of the template in a bidirectional consistency manner. The matching process uses a nearest neighbor search algorithm to find the best matching pair for each key point in the feature space. Through the iterative nearest point framework, the spatial transformation relationship between the two sets of key points is calculated, and the optimal affine transformation parameters are solved. This process is realized by minimizing the position error between all matching key point pairs, and the transformation parameters are updated after each iteration until the change in position error is less than 0.01 pixels.

[0080] The obtained optimal affine transformation parameters are used to perform global geometric transformation on the initial image set. The transformation parameters include two translation parameters, two scaling parameters and one rotation parameter. The pixel coordinates of each initial image are mapped to a new coordinate space through the affine transformation formula, and the pixel values at the new coordinate positions are calculated using the bilinear interpolation method. Through this transformation process, the transformer component structure in the image is accurately aligned in space with the standard template, and the average registration error between the output reference image and the standard template is controlled within 2 pixels, providing a geometrically consistent image basis for subsequent analysis.

[0081] In S3, based on the pixel-level difference between the initial image set and the reference image, the displacement vector field representing the local deformation of the image is inversely derived, specifically including:

[0082] A dense pixel correspondence relationship is established between the reference image and the corresponding images of the initial image set. A sliding window-based similarity calculation method is used, with each pixel as the center and a 7x7 pixel neighborhood window. The normalized cross-correlation value of the reference image and the initial image in the corresponding window is calculated, which is realized by first calculating the covariance of the two windows and then dividing by the product of the respective standard deviations. The normalized cross-correlation value ranges from -1 to 1, and the higher the value, the higher the similarity. All pixel positions in the image are traversed to generate an initial similarity distribution map, which reflects the matching degree of the reference image and the initial image in each local region.

[0083] Anisotropic diffusion filtering is performed on the initial similarity distribution map. This filtering process adjusts the diffusion coefficient based on image gradient information. In flat areas with small gradients, a larger diffusion coefficient is used to enhance smoothing, and in edge areas with large gradients, a smaller diffusion coefficient is used to maintain clear boundaries. In implementation, the diffusion coefficient is set to be proportional to the negative exponential function of the gradient value, and when the gradient value exceeds the set threshold of 15, the diffusion coefficient rapidly decreases. The filtering process is performed through iteration, updating the similarity value of each pixel point at each iteration, and the number of iterations is set to 20. After this processing, the connectivity of high-similarity regions in the similarity distribution map is enhanced, and local extreme values caused by noise are effectively suppressed, forming an optimized similarity distribution map.

[0084] Based on the optimized similarity distribution map, the maximum similarity path is tracked in the initial image from each pixel position in the reference image. The tracking process uses a gradient ascent algorithm, starting from each pixel position in the reference image, moving along the gradient direction of the similarity distribution map until the local maximum similarity point is found. The moving step size is dynamically adjusted according to the gradient size at the current position, using a smaller step size in the area with larger gradient to improve accuracy, and using a larger step size in the area with smaller gradient to speed up convergence. For each pixel, the spatial displacement vector from the starting point to the end point is recorded, which contains the horizontal and vertical displacement components. During the tracking process, the maximum number of iterations is set to 50, and when the distance difference between two consecutive movements is less than 0.1 pixels, it is considered that the maximum similarity point has been converged.

[0085] Finally, the spatial displacement vectors of all pixels are reorganized according to their original positions. A two-dimensional array with the same size as the reference image is established, and each element of the array corresponds to the displacement vector of a pixel. For each pixel position in the reference image, the calculated spatial displacement vector is stored in the array element corresponding to the position. Through such a reorganization process, a complete displacement vector field is constructed. The displacement vector field accurately represents the local geometric deformation of the initial image relative to the reference image, providing accurate deformation information for subsequent geometric correction. Each vector in the displacement vector field contains size and direction information, which can describe the deformation characteristics of each position in the image in detail.

[0086] In S4, the displacement vector field is smoothed and optimized under the geometric constraints of the transformer structure, and the initial image is geometrically reconstructed based on the optimized displacement field to generate a structure-true corrected image, specifically including:

[0087] Based on the material properties of each component of the transformer, the anisotropic elastic parameter field of the non-rigid connection area is established. The parameter field is set according to the actual structure characteristics of the transformer, where the heat sink connection area uses a higher elastic modulus parameter, with a value of 2.5 megapascals, and the cable joint area uses a lower elastic modulus parameter, with a value of 0.8 megapascals. The establishment process of the elastic parameter field uses a distance-weighted interpolation method to smoothly transition adjacent areas based on known component parameters. For each pixel position, according to its belonging to the structure area type, the corresponding elastic parameter value is assigned to form an elastic parameter distribution map covering the entire non-rigid connection area.

[0088] Firstly, the constitutive relationship between stress tensor and strain tensor is established at each pixel point, and the strain tensor is obtained by calculating the first-order spatial derivative of the displacement vector field. Specifically, at each pixel point, a 3x3 neighborhood window is taken, and the partial derivatives of the displacement field in the horizontal and vertical directions are calculated by the central difference method, and then the strain tensor components at this point are obtained. According to Hooke's law, the strain tensor is combined with the elastic parameter field to calculate the stress tensor at each pixel point. Substitute the stress tensor into the balance equation to get the partial differential equation about the second-order spatial derivative of the displacement vector field, which expresses the balance relationship between the internal stress of the elastic body and the external constraint.

[0089] The second-order derivative term in the partial differential equation is discretized by using the central difference format, and the displacement vector of each pixel point is linearly dependent on the displacement vectors of its adjacent 8 pixel points. In the discretization process, the grid spacing is taken as 1 pixel unit, and the difference coefficient is adjusted according to the spatial distribution of the elastic parameter field. Traverse all the pixel points in the non-rigid connection area, establish the corresponding discrete equation for each pixel point, arrange these equations according to the spatial position order of the pixel points, and assemble them into a large sparse linear equation system containing all unknown displacement vectors. The coefficient matrix of this equation system has the characteristics of symmetry and positive definiteness, and the non-zero elements are distributed on the main diagonal and its adjacent specific positions.

[0090] A hierarchical structure containing 3 grid levels is established, the grid spacing of the coarsest layer is 8 pixel units, the middle layer is 4 pixel units, and the finest layer is 1 pixel unit. In the coarse grid layer, the Gauss-Seidel iteration method is used to quickly eliminate the low-frequency components of the error, and the iteration number is set to 10 times; in the fine grid layer, the conjugate gradient method is used to accurately correct the high-frequency components of the error, and the iteration number is set to 30 times. Through the restriction and continuation operation between different grid layers, the effective transmission and elimination of error components are realized. The solving process is to make the residual norm less than As the convergence criterion, the displacement field correction is obtained which satisfies the mechanical equilibrium condition.

[0091] The displacement field correction is weighted and fused with the original displacement vector field. The weight factor is determined according to the confidence of each pixel point, which is calculated by the image texture richness and feature matching quality in the surrounding area of the point. For the area with rich texture and high feature matching quality, the original displacement field is given a larger weight, with a value between 0.7 and 0.9; for the area with sparse texture or poor feature matching quality, the displacement field correction is given a larger weight, with a value between 0.6 and 0.8. Through weighted average calculation, a smooth displacement field is generated which not only maintains the geometric deformation characteristics but also satisfies the mechanical continuity of the transformer structure.

[0092] The initial image is geometrically reconstructed based on the optimized smooth displacement field. A dense coordinate mapping relationship from the initial image to the reference image is established, which is determined by the smooth displacement field. An adaptive sampling algorithm is used to resample the initial image, using a cubic spline interpolation method in areas with large displacement gradients and a bilinear interpolation method in areas with small displacement gradients. During the resampling process, the sampling interval is set to 0.5 pixel units to ensure that the reconstructed image retains clear texture details. Through the coordinate mapping and resampling process, the geometric distortion in the initial image is eliminated, and a structure-corrected image is generated.

[0093] The accuracy of the correction process is ensured by verifying the consistency of the geometric dimensions of the key components of the transformer in the reconstructed image with the standard template. The key dimensions such as the diameter of the bushing and the spacing between the fins are measured and compared with the corresponding dimensions of the standard template, with the dimensional deviation controlled within 2%. At the same time, the outline integrity of the transformer components in the corrected image and the correctness of the connection relationship are checked to ensure that the generated corrected image conforms to the true geometric characteristics of the transformer structure, providing a geometric standardization analysis basis for subsequent fault feature extraction.

[0094] In S5, the fault feature is strengthened in the gradient domain of the corrected image, the abnormal region feature is extracted, the fault region is identified according to the abnormal region feature, and the final transformer fault diagnosis report is generated, which specifically includes:

[0095] In the corrected image after gradient domain strengthening, multi-channel feature extraction is performed on each candidate abnormal region. The extracted features include the hot spot distribution topological structure, the texture spectrum distribution entropy value, and the boundary tortuosity quantization parameter. The hot spot distribution topological structure is extracted by calculating the spatial distribution characteristics of high-temperature pixel points in the region, using a connectivity analysis method based on distance map to divide the hot spot region into different topological types, including isolated point distribution, linear distribution, and cluster distribution, etc. The texture spectrum distribution entropy value is calculated by analyzing the statistical characteristics of the gray level co-occurrence matrix in the region, selecting three texture features of contrast, correlation, and homogeneity, and calculating their information entropy values. The boundary tortuosity quantization parameter is obtained by calculating the curvature variation characteristics of the boundary curve, using a boundary description method based on chain code representation to count the frequency and amplitude of boundary direction changes.

[0096] The feature vector of the candidate abnormal region is constructed based on multi-channel features, and a reference fault feature vector with the same feature dimension is searched in a transformer typical fault feature library. The construction process of the feature vector quantizes the hot spot distribution topology structure into 3-dimensional features, quantizes the texture spectrum distribution entropy value into 3-dimensional features, and quantizes the boundary tortuosity parameter into 2-dimensional features, which together constitute an 8-dimensional feature vector. The transformer typical fault feature library contains standard feature vectors of common fault types of the transformer, including 10 typical fault types such as bushing overheating, winding deformation, and insulation deterioration, and 20 groups of standard feature vector samples are stored for each fault type.

[0097] The Mahalanobis distance of the candidate feature vector and each reference fault feature vector in the feature space is calculated. The calculation of the Mahalanobis distance needs to estimate the covariance matrix of all samples in the feature library first. By calculating the covariance values between each feature dimension, an 8x8 covariance matrix is constructed. Then the inverse matrix of the covariance matrix is calculated for standardizing the distance measurement in the feature space. In the specific calculation, the difference vector is obtained by subtracting each reference feature vector from the candidate feature vector, and then the difference vector is multiplied by the inverse covariance matrix, and then multiplied by the transpose of the difference vector, and finally the square root is taken to obtain the Mahalanobis distance value. This calculation process fully considers the correlation and dimension difference between different feature dimensions.

[0098] The first three reference fault types with the smallest Mahalanobis distance are selected as the preliminary results, and the corresponding distance values are recorded as the matching degree indicators. The matching degree indicators are calculated by normalizing the Mahalanobis distance values, and using a negative exponential function to convert the distance values to matching degree scores between 0 and 1. The smaller the distance, the higher the matching degree score. When selecting the preliminary results, the absolute threshold of the Mahalanobis distance is set to 5.0, and the fault types exceeding the threshold are not considered to ensure the reliability of the preliminary results.

[0099] Based on the Mahalanobis distance distribution characteristics of the three preliminary results, the confidence level of each preliminary result is calculated through distance ratio analysis and relative position evaluation. The distance ratio analysis includes calculating the ratio of the smallest Mahalanobis distance to the second smallest Mahalanobis distance, and the ratio of the smallest distance to the third smallest distance. When the ratio of the smallest distance to the second smallest distance is less than 0.7, it is considered that the confidence of the primary candidate result is high; when the ratio is greater than 0.9, it is considered that the discrimination of the first two candidate results is insufficient. The relative position evaluation analyzes the distribution pattern by calculating the area and shape features of the triangle formed by the three candidate results in the feature space. The confidence level is calculated by combining the distance ratio and the relative position features, and the weighted sum method is used, in which the weight of the distance ratio is set to 0.6 and the weight of the relative position is set to 0.4.

[0100] An adaptive morphological closing operation is performed on the candidate regions based on the primary election results to fuse broken boundaries and fill internal holes. The size of the structural element of the morphological closing operation is adaptively determined according to the geometric characteristics of the fault region. For a linear fault region, a 3x1 rectangular structural element is used, and for a planar fault region, a 3x3 circular structural element is used. The operation process first performs dilation operation to connect the broken boundary fragments, and then performs erosion operation to restore the original size of the region. Through this processing, a complete three-dimensional geometric representation of the fault region is generated, including geometric parameters such as area, perimeter, and minimum bounding rectangle.

[0101] The final diagnosis report is automatically generated according to the preset diagnosis report template based on the fault type, confidence level, geometric parameters, and spatial position information. The report template includes mandatory fields such as fault type description, confidence level, fault region size, and spatial position coordinates. The fault type description lists the three candidate results from high to low according to the confidence level, and marks the confidence values of each result. The geometric parameters include the area, maximum size, and shape characteristics of the fault region, with the area unit converted to square centimeters and the size unit converted to centimeters. The spatial position coordinates are expressed in relative coordinates relative to the transformer reference point, including horizontal distance and vertical height. The diagnosis report also includes the severity level assessment of the fault, which is automatically divided into three levels of slight, moderate, and severe according to the fault region size and type, and provides corresponding treatment suggestions for each level. The final diagnosis report is output in the form of a structured document, including text description, data table, and labeled schematic diagram, ensuring the completeness and readability of the diagnosis results.

[0102] Referring to Figure 2 As shown in the figure, the transformer fault diagnosis system based on image recognition includes:

[0103] The dynamic sequence acquisition and quality screening module is used to acquire a multi-modal image sequence of the transformer component in real time, and automatically screen out effective images containing key fault features based on image clarity and component integrity evaluation to form an initial image set;

[0104] The multi-scale template registration module is used to perform multi-scale spatial registration between the initial image set and the pre-constructed transformer normal state standard template to generate a structure-aligned reference image;

[0105] The deformation field reverse derivation module reversely derives a displacement vector field representing local deformation of the image based on the pixel-level difference between the initial image set and the reference image;

[0106] The physical constraint geometry correction module is used to smooth and optimize the displacement vector field under the geometric constraints of the transformer structure, and perform geometric reconstruction on the initial image based on the optimized displacement field to generate a structure-true corrected image;

[0107] A fault diagnosis and report generation module is configured to enhance fault features in the gradient domain of the corrected image, extract abnormal region features, identify a fault region based on the abnormal region features, and generate a final transformer fault diagnosis report.

[0108] The working principle of the present application is as follows: precise fault detection is achieved through multi-stage image processing and feature analysis. Multi-modal image sequences of transformer components are synchronously collected by deployed visible light cameras and infrared thermal imagers, and an initial image set is formed after quality screening based on definition and component integrity. Then, the initial image set is multi-scale spatially registered with a normal state standard template of the transformer to generate a reference image, and a displacement vector field is inversely deduced based on pixel-level differences. Then, the displacement vector field is smoothed and optimized under the geometric constraints of the transformer structure, and the initial image is geometrically reconstructed based on the optimized displacement field to obtain a corrected image. Finally, fault features are enhanced in the gradient domain of the corrected image, and the fault region is identified through multi-channel feature extraction and matching with a typical transformer fault feature library. A complete diagnosis report is generated by integrating fault type, confidence, geometric parameters and spatial location information. The method realizes accurate identification and quantitative analysis of transformer faults through a complete image processing chain.

[0109] The above describes one embodiment of the present application in detail, but the content described is only a preferred embodiment of the present application and cannot be considered as limiting the scope of the present application. Any equivalent changes and improvements made in accordance with the scope of the present application should still belong to the scope of the present application.

Claims

1. A transformer fault diagnosis method based on image recognition, characterized by, The method comprises the following steps: S1: Real-time acquisition of a multi-modal image sequence of the transformer component, and automatic screening of effective images containing key fault features based on image definition and component integrity evaluation to form an initial image set; S2: Multi-scale spatial registration of the initial image set with a pre-constructed transformer normal state standard template to generate a structure-aligned reference image; The generated structure-aligned reference image specifically comprises: Extracting the transformer component dissection semantic key points of each image from the initial image set, the key points including sleeve flange edge joint points, heat sink rib plate root connection points, and oil tank weld feature intersection points; Constructing a Gaussian difference pyramid representation of the transformer normal state standard template, and establishing a scale-invariant spatial coordinate set of the dissection semantic key points at each pyramid level; Bidirectional consistency matching of the dissection semantic key points and the scale-invariant spatial coordinate set, and solving optimal affine transformation parameters through an iterative closest point framework; Global geometric transformation of the initial image set by using the solved optimal affine transformation parameters, so that the transformer component structure in the image is accurately aligned with the standard template in space, and a structure-aligned reference image is output; S3: Reverse derivation of a displacement vector field representing local deformation of the image based on pixel-level differences between the initial image set and the reference image; The reverse derivation of the displacement vector field representing local deformation of the image specifically comprises: Establishing a dense pixel correspondence relationship between the reference image and the corresponding image of the initial image set, quantifying the local similarity of the image by calculating the normalized cross-correlation value in the neighborhood of each pixel, and generating an initial similarity distribution map; Anisotropic diffusion filtering of the initial similarity distribution map, enhancing the connectivity of high-similarity areas along the image gradient direction, and suppressing local extrema caused by noise to form an optimized similarity distribution map; Based on the optimized similarity distribution map, the spatial offset vector of each pixel to the corresponding maximum similarity point is calculated by tracking the maximum similarity path in the initial image from the starting point of each pixel position in the reference image; Reorganizing the spatial offset vectors of all pixels according to their original positions to construct a complete displacement vector field; S4: Smoothing optimization of the displacement vector field under the geometric constraint of the transformer structure, and geometric reconstruction of the initial image based on the optimized displacement field to generate a structure-true corrected image; S5: Fault feature enhancement in the gradient domain of the corrected image, extraction of abnormal region features, identification of fault regions according to the abnormal region features, and generation of a final transformer fault diagnosis report.

2. The image recognition-based transformer fault diagnosis method according to claim 1, characterized in that, The initial image set is specifically formed as follows: Synchronously triggering the visible light image sequence and the infrared image sequence of the transformer component by deploying fixed visible light cameras and infrared thermographs around the transformer to form a raw multi-modal image data stream; For each frame of image in the raw multi-modal image data stream, the image definition is evaluated by using variance calculation based on second-order differential features, and the frames meeting the definition are screened out by comparing with a preset dynamic threshold to generate a preliminary screening image set; Based on the prior three-dimensional geometric information of the transformer components, a standard contour projection template of the key components is constructed, each image in the preliminary screening image set is spatially registered with the template, the overlapping ratio of the visible area of the components is calculated as the integrity score, and images with scores lower than the integrity threshold are removed; Multi-scale gradient domain analysis is performed on the images that pass the integrity evaluation, potential fault features are preliminarily identified by extracting high-frequency gradient features and statistically analyzing their distribution abnormalities, and finally these images are collected as effective images to form an initial image set.

3. The image recognition-based transformer fault diagnosis method according to claim 1, characterized in that, The method for generating a corrected image with anatomical reality comprises the following steps: Identify the areas corresponding to the rigid structural components of the transformer, including the flange of the bushing and the wall joint of the oil tank, and apply rigid motion constraints to these areas to maintain their geometric invariance; Establish structural continuity constraints for the areas corresponding to the non-rigid connectors in the displacement vector field, and obtain a smooth displacement field that satisfies the physical structure characteristics of the transformer by solving a system of partial differential equations that satisfy the equilibrium conditions of elastic mechanics; Based on the smoothed displacement field, a dense coordinate mapping relationship between the initial image and the reference image is established, and the initial image is resampled to eliminate geometric distortion while maintaining the integrity of the texture features; By verifying the consistency of the geometric dimensions of the key components of the transformer in the reconstructed image with the standard template, a corrected image that conforms to the anatomical reality is generated.

4. The image recognition-based transformer fault diagnosis method according to claim 3, characterized in that, The method for generating a corrected image with anatomical reality comprises the following steps: Based on the material properties of each component of the transformer, an anisotropic elastic parameter field of the non-rigid connection area is established; In the non-rigid connection area of the displacement vector field, a balance equation based on the stress-strain relationship is constructed, and the continuous partial differential equation is discretized into a linear equation system by the finite difference method; Solve the linear equation system to quickly eliminate the low-frequency components of the error in the coarse grid layer and accurately correct the high-frequency components of the error in the fine grid layer to obtain a displacement field correction that satisfies the mechanical equilibrium conditions; The displacement field correction and the original displacement vector field are fused by weighting to generate a smooth displacement field that maintains the geometric deformation characteristics and satisfies the structural mechanics continuity of the transformer.

5. The image recognition-based transformer fault diagnosis method according to claim 4, characterized in that, The method for generating a corrected image with anatomical reality comprises the following steps: Based on the anisotropic elastic parameter field, a constitutive relationship between the stress tensor and the strain tensor is established at each pixel point in the non-rigid connection area, where the strain tensor is determined by the first-order spatial differentiation of the displacement vector field; Substitute the stress tensor of each pixel point into the balance equation to obtain a partial differential equation about the second-order spatial differentiation of the displacement vector field; Discretize the second-order differentiation term in the partial differential equation to establish a linear dependence relationship between the displacement vector of each pixel point and the displacement vectors of its neighboring pixel points; Traverse all pixel points in the non-rigid connection area, and assemble the discretized equations into a large sparse linear equation system containing all unknown displacement vectors according to the spatial position order.

6. The image recognition-based transformer fault diagnosis method according to claim 1, characterized in that, The illustrated fault region is identified according to the abnormal region feature, and a final transformer fault diagnosis report is generated, specifically comprising: In the corrected image after gradient domain reinforcement, multi-channel feature extraction is performed on each candidate abnormal region, and the multi-channel feature includes the hot spot distribution topological structure, texture spectrum distribution entropy value and boundary tortuosity quantitative parameter of the region; The extracted multi-channel feature is bidirectionally matched with the transformer typical fault feature library, the Mahalanobis distance in the feature space is calculated to determine the fault type preliminary selection result, and the confidence level is evaluated; Based on the preliminary selection result, adaptive morphological closing operation is performed on the candidate region, the broken boundary is fused and the internal hole is filled, and the complete fault region three-dimensional geometric representation is generated; Comprehensive fault type, confidence, geometric parameter and spatial position information, according to the preset diagnosis report template, the final diagnosis report containing fault qualitative conclusion and quantitative analysis data is automatically generated.

7. The image recognition-based transformer fault diagnosis method according to claim 6, characterized in that, The extracted multi-channel feature is bidirectionally matched with the transformer typical fault feature library, the Mahalanobis distance in the feature space is calculated to determine the fault type preliminary selection result, and the confidence level is evaluated, specifically comprising: Based on the multi-channel feature, a feature vector of the candidate abnormal region is constructed, and a reference fault feature vector with the same feature dimension is searched in the transformer typical fault feature library; The Mahalanobis distance of the candidate feature vector and each reference fault feature vector in the feature space is calculated, and the Mahalanobis distance is obtained according to the covariance relationship between different feature dimensions; The first three reference fault types with the smallest Mahalanobis distance are selected as the preliminary selection result, and the corresponding distance value is recorded as the matching degree index; Based on the Mahalanobis distance distribution characteristics of the three preliminary selection results, the confidence level of each preliminary selection result is calculated through distance ratio analysis and relative position evaluation, and a fault type preliminary selection result set with confidence weight is formed.

8. A transformer fault diagnosis system based on image recognition, characterized by, The image recognition-based transformer fault diagnosis method of any one of claims 1-7, comprising: A dynamic sequence acquisition and quality screening module is used to acquire a multi-modal image sequence of the transformer component in real time, and automatically screen out effective images containing key fault features based on image clarity and component integrity evaluation to form an initial image set; A multi-scale template registration module is used to perform multi-scale space registration of the initial image set and the pre-constructed transformer normal state standard template to generate a structure-aligned reference image; A deformation field reverse derivation module is used to reversely derive a displacement vector field representing local deformation of the image based on the pixel-level difference between the initial image set and the reference image; A physical constraint geometric correction module is used to smooth and optimize the displacement vector field under the geometric constraint of the transformer structure, and perform geometric reconstruction on the initial image based on the optimized displacement field to generate a structure-true corrected image; A fault diagnosis and report generation module is used to perform fault feature reinforcement in the gradient domain of the corrected image, extract abnormal region features, identify fault regions according to abnormal region features, and generate a final transformer fault diagnosis report.

Images