Transformer fault diagnosis method and system based on image recognition

Abstract

The invention relates to the technical field of power equipment state monitoring, and particularly discloses a transformer fault diagnosis method and system based on image recognition, and the method comprises the steps: collecting a transformer multi-mode image sequence in real time, and carrying out the definition and part integrity evaluation and screening 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 reversely deriving a displacement vector field based on pixel-level difference; carrying out smooth optimization and geometric reconstruction on the displacement vector field under the geometric constraint of the transformer structure, and generating a correction image with a real structure; fault feature enhancement is carried out in a gradient domain of the corrected image, a fault area is identified through matching of multichannel feature extraction and a transformer typical fault feature library, and a diagnosis report integrating fault types, confidence coefficients and geometric parameters is generated; according to the method, the problem of image geometric deformation caused by shooting condition differences is effectively solved, and the accuracy and reliability of fault identification are improved.

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 manual periodic 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 structure-corrected image, which fundamentally eliminates the geometric errors introduced by the acquisition angle and distance, and establishes a stable and reliable image basis for the 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 above background.

[0005] The purpose of the present application can be achieved by the following technical solutions: The transformer fault diagnosis method based on image recognition comprises the following steps: 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; 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; 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; S4: Smoothing optimization of the displacement vector field under the geometric constraints of the transformer structure, and geometric reconstruction of the initial image based on the optimized displacement field to generate a structure-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.

[0006] As a further aspect of the present application: the initial image set is specifically composed of: By deploying fixed visible light cameras and infrared thermographs around the transformer, visible light image sequences and infrared image sequences of the transformer components are collected in a synchronous triggering manner to form a raw multi-modal image data stream; For each frame of image in the raw multi-modal image data stream, the image sharpness is evaluated by using variance calculation based on second-order differential features, and the frames with qualified sharpness are selected 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, and the overlap ratio of the visible region of the component is calculated as the integrity score, and the images with a score 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 the initial image set.

[0007] As a further aspect of the present application: the generation of the structure-aligned reference image specifically includes: The anatomical semantic key points of the transformer components are extracted from each image in the initial image set, including the sleeve flange edge joint, the heat sink rib root connection point, and the oil tank weld feature intersection point; 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; The anatomical semantic key points are matched with the scale-invariant spatial coordinate set in a bidirectional consistency manner, and the optimal affine transformation parameters are solved by using the iterative closest point framework; The optimal affine transformation parameters obtained are used to perform global geometric transformation on the initial image set, so that the structure of the transformer components in the image is accurately aligned with the standard template in space, and the structure-aligned reference image is output.

[0008] As a further aspect of the present application: the displacement vector field representing the local deformation of the image is reversely derived, specifically including: A dense pixel correspondence relationship is established between the reference image and the corresponding image in the initial image set, the local similarity of the image is quantified by calculating the normalized cross-correlation value in the neighborhood of each pixel, and an initial similarity distribution map is generated; Anisotropic diffusion filtering is performed on the initial similarity distribution map to enhance the connectivity of high-similarity regions along the image gradient direction and suppress local extreme values caused by noise, thereby forming an optimized similarity distribution map; Based on the optimized similarity distribution map, a spatial displacement vector of each pixel to the corresponding maximum similarity point is calculated by tracking the maximum similarity path in the initial image starting from each pixel position in the reference image; The spatial displacement vectors of all pixels are reorganized according to their original positions to construct a complete displacement vector field.

[0009] As a further scheme of the present application, the generated structure-true corrected image specifically comprises: Regions corresponding to rigid structure components of the transformer, including the sleeve flange and the oil tank wall joint, are identified, and rigid motion constraints are applied to these regions to maintain their geometric invariance; Regions corresponding to non-rigid connectors in the displacement vector field are established with structural continuity constraints, and 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; Based on the smoothed displacement field, a dense coordinate mapping relationship from the initial image to 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 conforming to the anatomical reality is generated.

[0010] 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, specifically comprising: Anisotropic elastic parameter fields of non-rigid connection regions are established based on the material properties of each component of the transformer; Balance equations based on stress-strain relationships are constructed in the non-rigid connection regions of the displacement vector field, and the continuous partial differential equations are discretized into linear equation groups by the finite difference method; The linear equation groups are solved to quickly eliminate the low-frequency components of errors in the coarse grid layer and accurately correct the high-frequency components of errors in the fine grid layer, thereby obtaining a displacement field correction that satisfies the mechanical equilibrium conditions; The displacement field correction and the original displacement vector field are weighted and fused to generate a smooth displacement field that maintains the geometric deformation characteristics and satisfies the structural mechanics continuity of the transformer.

[0011] As a further aspect of the present invention: the construction of equilibrium equations based on stress-strain relationships in the non-rigidly connected region of the displacement vector field, and the discretization of continuous partial differential equations into a system of linear equations using the finite difference method, specifically includes: Based on the anisotropic elastic parameter field, a constitutive relationship between the stress tensor and the strain tensor is established at each pixel in the non-rigid connection region, where the strain tensor is determined by the first spatial differential of the displacement vector field. Substituting the stress tensor of each pixel into the equilibrium equation, we obtain a partial differential equation for the second-order spatial differential of the displacement vector field. Discretize the second-order differential terms in the partial differential equation and establish a linear dependency between the displacement vector of each pixel and the displacement vectors of its neighboring pixels. Traverse all pixels in the non-rigidly connected region and assemble the discretized equations into a large sparse linear system of equations containing all unknown displacement vectors in spatial order.

[0012] As a further aspect of the present invention: the method of identifying fault areas based on abnormal area characteristics and generating a final transformer fault diagnosis report specifically includes: In the corrected image enhanced in the gradient domain, multi-channel features are extracted for each candidate anomalous region. The multi-channel features include the hot spot distribution topology, texture spectrum distribution entropy, and boundary tortuosity quantification parameters of the region. The extracted multi-channel features are bidirectionally matched with the typical transformer fault feature library. The initial fault type selection results are determined by calculating the Mahalanobis distance in the feature space, and the confidence level is evaluated. Based on the preliminary selection results, an adaptive morphological closing operation is performed on the candidate region to fuse the fracture boundary and fill the internal holes, generating a complete three-dimensional geometric representation of the fault region. Based on the comprehensive fault type, confidence level, geometric parameters, and spatial location information, the system automatically generates a final diagnostic report containing qualitative conclusions and quantitative analysis data according to a preset diagnostic report template.

[0013] As a further aspect of the present invention: the step of bidirectionally matching the extracted multi-channel features with a typical transformer fault feature library, determining the initial fault type selection result by calculating the Mahalanobis distance in the feature space, and evaluating its confidence level specifically includes: Feature vectors of candidate abnormal regions are constructed based on multi-channel features, and reference fault feature vectors with the same feature dimensions are retrieved from the typical transformer fault feature library. Calculate the Mahalanobis distance between the candidate feature vector and each reference fault feature vector in the feature space. The Mahalanobis distance is obtained based on the covariance relationship between different feature dimensions. The three reference fault types with the smallest Mahalanobis distance were selected as the initial selection results, and the corresponding distance values ​​were recorded as matching degree indicators. 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, forming a set of preliminary selection results for fault types with confidence weights.

[0014] The image recognition-based transformer fault diagnosis system includes: The dynamic sequence acquisition and quality screening module is used to acquire multimodal image sequences of transformer components in real time, and automatically screen out valid images containing key fault features based on image clarity and component integrity assessment to form an initial image set; The multi-scale template registration module is used to perform multi-scale spatial registration between the initial image set and the pre-constructed standard template of the transformer normal state to generate a reference image with structural alignment. The deformation field inverse derivation module, based on the pixel-level difference between the initial image set and the reference image, inversely derives the displacement vector field characterizing the local deformation of the image; The physical constraint geometric correction module is used to smooth and optimize the displacement vector field under the geometric constraints of the transformer structure, and to perform geometric reconstruction of the initial image based on the optimized displacement field to generate a corrected image of the structure that is true to the structure. The fault diagnosis and report generation module is used to enhance fault features in the gradient domain of the corrected image, extract abnormal region features, identify fault regions based on abnormal region features, and generate the final transformer fault diagnosis report.

[0015] The beneficial effects of this invention are: (1) By establishing a standard template for normal transformer status and a multi-scale spatial registration method, combined with displacement vector field derivation and geometric reconstruction under physical constraints, the geometric distortion of images caused by changes in shooting angle and distance is effectively eliminated, making 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 geometrically standardized analytical basis for fault diagnosis.

[0016] (2) By adopting multimodal image sequence analysis and gradient domain fault feature enhancement technology, through multi-channel feature extraction such as hot spot distribution topology and texture spectrum distribution entropy value, combined with feature matching and confidence evaluation mechanism based on Mahalanobis distance, sensitive detection and accurate identification of early fault features of transformers can be achieved. It can effectively distinguish between normal state and potential fault, and improve the accuracy and reliability of fault diagnosis. Attached Figure Description

[0017] The invention will now be further described with reference to the accompanying drawings.

[0018] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a system block diagram of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Please see Figure 1 As shown, this invention is a transformer fault diagnosis method based on image recognition, comprising the following steps: S1: Real-time acquisition of multimodal image sequences of transformer components, and automatic filtering of valid images containing key fault features based on image clarity and component integrity assessment to form an initial image set; S2: Perform multi-scale spatial registration between the initial image set and the pre-constructed standard template for normal transformer states to generate a reference image with structural alignment; S3: Based on the pixel-level differences between the initial image set and the reference image, the displacement vector field characterizing the local deformation of the image is derived in reverse. S4: Under the geometric constraints 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 corrected image of the structure. S5: Perform fault feature enhancement in the gradient domain of the corrected image, extract abnormal region features, identify fault regions based on abnormal region features, and generate the final transformer fault diagnosis report.

[0021] In S1, multimodal image sequences of transformer components are acquired in real time, and valid images containing key fault features are automatically selected based on image clarity and component integrity assessment to form an initial image set, specifically including: Fixed visible light cameras and infrared thermal imagers deployed around the transformer are used to acquire visible light and infrared image sequences of transformer components in a synchronously triggered manner. The visible light cameras capture detailed information about the transformer surface condition, while the infrared thermal imagers simultaneously acquire temperature distribution information of the same areas, forming a raw multimodal image data stream aligned in time and space.

[0022] Sharpness assessment is performed on each frame of the original multimodal image data stream. A sharpness evaluation method based on the Laplacian operator is adopted, which performs second-order differential processing on the image and calculates the variance of the processed image as the sharpness evaluation value. A dynamic threshold adjustment mechanism is set, which adaptively adjusts the threshold according to ambient lighting conditions and shooting distance. Image frames with sharpness evaluation values ​​below the threshold are removed, and image frames that meet the sharpness standard are retained to generate a preliminary selected image set.

[0023] Based on the prior 3D geometric information of transformer components, standard contour projection templates for key components are constructed. These templates include the standard outlines of major components such as transformer bushings, heat sinks, and oil tanks. Each image in the initially selected image set is spatially registered with the corresponding standard contour projection template. The overlap ratio between the visible area of ​​the component in the actual captured image and the template outline is calculated, and this ratio is used as the component integrity score. An integrity threshold of 85% is set, and images with integrity scores below this threshold are discarded.

[0024] Multi-scale gradient domain analysis was performed on images that passed the integrity assessment. A Gaussian pyramid was constructed to extract high-frequency gradient features at three different scales. The distribution characteristics of these gradient features at each scale were statistically analyzed, and the skewness and kurtosis of the gradient distribution were calculated as anomaly assessment indicators. When the absolute value of the gradient distribution skewness was greater than 0.5 or the kurtosis exceeded the range of 3.2-4.8, the image was determined to contain potential fault features. All images determined to contain potential fault features were compiled to form the initial image set for subsequent analysis.

[0025] In S2, the initial image set is spatially registered with a pre-constructed standard template for the normal state of a transformer at multiple scales to generate a structurally aligned reference image, specifically including: Anatomical semantic key points of transformer components were extracted from each image in the initial image set. These key points include bushing flange edge joints, radiator fin root connection points, and weld feature intersections in the tank. The key point extraction process employed a gradient orientation histogram-based feature detection method, which identifies key point locations with stable geometric characteristics by calculating the statistical features of gradient orientation distribution within local image regions. For each candidate key point, its response value in scale space was calculated, and stable key points with response values ​​exceeding a preset threshold were retained, ultimately forming the anatomical semantic key point set for each image.

[0026] A Gaussian difference pyramid representation of a standard template for a transformer's normal state is constructed. A pyramid structure with five scale levels is established, where the image at each scale is convolved using a Gaussian kernel function, with the scale factor increasing in multiples of 1.2. At each level of the Gaussian difference pyramid, a scale-invariant set of spatial coordinates for anatomical semantic keypoints is established based on the geometric characteristics of the template image. These coordinate sets are normalized using the relative positional relationships of the keypoints at different scales to ensure scale invariance.

[0027] The anatomical semantic keypoints extracted from the initial image are bidirectionally matched with the scale-invariant spatial coordinate set of the template. The matching process employs a nearest neighbor search algorithm to find the best matching pair for each keypoint in the feature space. By iteratively applying the nearest point framework, the spatial transformation relationship between the two sets of keypoints is calculated, and the optimal affine transformation parameters are solved. This process is achieved by minimizing the positional error between all matching keypoint pairs; the transformation parameters are updated after each iteration until the change in positional error is less than 0.01 pixels.

[0028] A global geometric transformation is performed on the initial image set using the optimal affine transformation parameters obtained from the solution. 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 using the affine transformation formula, and the pixel values ​​at the new coordinate positions are calculated using bilinear interpolation. This transformation process ensures that the transformer component structure in the image is precisely aligned spatially with the standard template. The average registration error between the output reference image and the standard template is controlled within 2 pixels, providing a geometrically consistent image foundation for subsequent analysis.

[0029] In S3, based on the pixel-level differences between the initial image set and the reference image, the displacement vector field characterizing the local deformation of the image is derived in reverse, specifically including: A dense pixel correspondence is established between the reference image and corresponding images in the initial image set. A sliding window-based similarity calculation method is used, with a 7×7 pixel neighborhood window centered on each pixel. The normalized cross-correlation value between the reference image and the initial image within the corresponding window is calculated. This calculation is achieved by first calculating the covariance of the two windows and then dividing by the product of their respective standard deviations. The normalized cross-correlation value ranges from -1 to +1, with a higher value indicating higher similarity. All pixel positions in the image are traversed to generate an initial similarity distribution map, which reflects the degree of matching between the reference image and the initial image in each local region.

[0030] Anisotropic diffusion filtering is applied to the initial similarity distribution map. This filtering process adjusts the diffusion coefficient based on image gradient information, using a larger diffusion coefficient in flat regions with smaller gradients to enhance smoothness, and a smaller diffusion coefficient in edge regions with larger gradients to maintain clear boundaries. Specifically, the diffusion coefficient is set to be proportional to a negative exponential function of the gradient value; when the gradient value exceeds a set threshold of 15, the diffusion coefficient decreases rapidly. The filtering process is performed iteratively, updating the similarity value of each pixel in each iteration, with 20 iterations. After this processing, the connectivity of high-similarity regions in the similarity distribution map is enhanced, while local maxima caused by noise are effectively suppressed, resulting in an optimized similarity distribution map.

[0031] Based on the optimized similarity distribution map, the maximum similarity path is traced in the initial image, starting from each pixel position in the reference image. The tracing process employs a gradient ascent algorithm, moving along the gradient direction of the similarity distribution map from each pixel position in the reference image until a local maximum similarity point is found. The step size is dynamically adjusted based on the gradient magnitude at the current position; a smaller step size is used in regions with larger gradients to improve accuracy, and a larger step size is used in regions with smaller gradients to accelerate convergence. For each pixel, a spatial offset vector from the starting point to the ending point is recorded, containing horizontal and vertical displacement components. During the tracing process, the maximum number of iterations is set to 50. Convergence to the maximum similarity point is considered achieved when the difference in distance between two consecutive moves is less than 0.1 pixels.

[0032] Finally, the spatial offset vectors of all pixels are reassembled according to their original positions. A two-dimensional array of the same size as the reference image is constructed, with each element of the array corresponding to the displacement vector of a pixel. For each pixel position in the reference image, its calculated spatial offset vector is stored in the corresponding array element. Through this reassembly process, a complete displacement vector field is constructed. This displacement vector field accurately characterizes the local geometric deformation of the initial image relative to the reference image, providing precise deformation information for subsequent geometric correction. Each vector in the displacement vector field contains magnitude and orientation information, which can describe in detail the deformation characteristics of each location in the image.

[0033] In S4, under the geometric constraints 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 corrected image that accurately reflects the structure. Specifically, this includes: An anisotropic elastic parameter field was established for the non-rigid connection region based on the material properties of each component of the transformer. This parameter field was set according to the actual structural characteristics of the transformer; a higher elastic modulus parameter, set to 2.5 MPa, was used for the heat sink connection region, while a lower elastic modulus parameter, set to 0.8 MPa, was used for the cable joint region. The establishment of the elastic parameter field employed a distance-weighted interpolation method, performing smooth transition processing on adjacent regions based on known component parameters. For each pixel location, a corresponding elastic parameter value was assigned according to its structural region type, forming an elastic parameter distribution map covering the entire non-rigid connection region.

[0034] First, constitutive relations between the stress tensor and strain tensor are established at each pixel. The strain tensor is obtained by calculating the first-order spatial differential of the displacement vector field. Specifically, at each pixel, a 3×3 neighborhood window is taken, and the partial derivatives of the displacement field in the horizontal and vertical directions are calculated using the central difference method, thus obtaining the strain tensor components at that point. According to Hooke's law, the strain tensor is combined with the elastic parameter field to calculate the stress tensor at each pixel. Substituting the stress tensor into the equilibrium equation, a partial differential equation for the second-order spatial differential of the displacement vector field is obtained. This equation expresses the equilibrium relationship between the internal stress and external constraints of the elastic body.

[0035] The second-order differential terms in the partial differential equations are discretized using a central difference scheme, establishing a linear dependency between the displacement vector of each pixel and the displacement vectors of its eight neighboring pixels. During discretization, the grid spacing is set to one pixel, and the difference coefficients are adjusted according to the spatial distribution of the elastic parameter field. All pixels in the non-rigidly connected region are traversed, and a corresponding discrete equation is established for each pixel. These equations are arranged according to the spatial position of the pixels, assembling into a large sparse linear system of equations containing all unknown displacement vectors. The coefficient matrix of this system of equations exhibits symmetric positive definiteness, with non-zero elements distributed along the main diagonal and in specific adjacent positions.

[0036] A hierarchical structure with three grid levels is established: the coarsest layer has a grid spacing of 8 pixels, the middle layer 4 pixels, and the finest layer 1 pixel. In the coarse grid layer, the Gauss-Seidel iterative method is used to quickly eliminate low-frequency components of the error, with 10 iterations. In the fine grid layer, the conjugate gradient method is used to accurately correct high-frequency components of the error, with 30 iterations. Through constraints and extension operations between different grid layers, effective propagation and elimination of error components are achieved. The solution process uses a residual norm less than... As a convergence criterion, the displacement field correction that satisfies the mechanical equilibrium condition is obtained.

[0037] The displacement field correction is weighted and fused with the original displacement vector field. The weighting factor is determined based on the confidence level of each pixel, which is calculated by the richness of image texture and the quality of feature matching in the surrounding area. Regions with rich texture and high feature matching quality are assigned a larger weight to the original displacement field, ranging from 0.7 to 0.9; regions with sparse texture or poor feature matching quality are assigned a larger weight to the displacement field correction, ranging from 0.6 to 0.8. Through weighted averaging, a smooth displacement field is generated that maintains both geometric deformation characteristics and satisfies the structural mechanical continuity of the transformer.

[0038] Geometric reconstruction of the initial image is performed based on the optimized smooth displacement field. A dense coordinate mapping relationship from the initial image to the reference image is established, determined by the smooth displacement field. An adaptive sampling algorithm is used to resample the initial image, employing cubic spline interpolation in regions with large displacement gradients and bilinear interpolation in regions with small displacement gradients. During resampling, a sampling interval of 0.5 pixels is set to ensure that the reconstructed image retains clear texture details. Through coordinate mapping and resampling, geometric distortions in the initial image are eliminated, generating a structurally accurate corrected image.

[0039] The accuracy of the calibration process is ensured by verifying the consistency of the geometric dimensions of key transformer components in the reconstructed image with the standard template. Key dimensions such as bushing diameter and heat sink spacing in the calibrated image are measured and compared with corresponding dimensions in the standard template, with dimensional deviations controlled within 2%. Simultaneously, the integrity of the transformer component outlines and the correctness of their connections in the calibrated image are checked to ensure that the generated calibrated image conforms to the true geometric characteristics of the transformer structure, providing a geometrically standardized analytical basis for subsequent fault feature extraction.

[0040] In S5, fault feature enhancement is performed within the gradient domain of the corrected image, abnormal region features are extracted, fault regions are identified based on these abnormal region features, and a final transformer fault diagnosis report is generated, specifically including: In the corrected image enhanced in the gradient domain, multi-channel feature extraction is performed on each candidate anomaly region. The extracted features include the hotspot distribution topology, texture spectrum distribution entropy, and boundary tortuosity quantification parameters. The hotspot distribution topology is extracted by calculating the spatial distribution characteristics of high-temperature pixels within the region. A connectivity analysis method based on distance maps is used to classify the hotspot region into different topological types, including isolated point distributions, linear distributions, and clumped distributions. The texture spectrum distribution entropy is calculated by analyzing the statistical characteristics of the gray-level co-occurrence matrix within the region, selecting contrast, correlation, and homogeneity as three texture features, and calculating their information entropy values. The boundary tortuosity quantification parameters are obtained by calculating the curvature change characteristics of the boundary curves. A boundary description method based on chain code representation is used to statistically analyze the frequency and amplitude of boundary direction changes.

[0041] Feature vectors for candidate anomaly regions are constructed based on multi-channel features, while reference fault feature vectors with the same feature dimensions are retrieved from the transformer typical fault feature library. The feature vector construction process quantizes the hotspot distribution topology into 3D features, the texture spectrum distribution entropy into 3D features, and the boundary tortuosity quantization parameter into 2D features, collectively forming an 8-dimensional feature vector. The transformer typical fault feature library contains standard feature vectors for common transformer fault types, including 10 typical fault types such as bushing overheating, winding deformation, and insulation degradation. Each fault type stores 20 sets of standard feature vector samples.

[0042] The Mahalanobis distance between candidate feature vectors and each reference fault feature vector in the feature space is calculated. The calculation of the Mahalanobis distance requires first estimating the covariance matrix of all samples in the feature library. An 8×8 covariance matrix is ​​constructed by calculating the covariance values ​​between each feature dimension. Then, the inverse of this covariance matrix is ​​calculated and used to standardize the distance metric in the feature space. Specifically, the candidate feature vector is subtracted from each reference feature vector to obtain a difference vector. This difference vector is multiplied by the inverse covariance matrix, 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 dimensional differences between different feature dimensions.

[0043] The three reference fault types with the smallest Mahalanobis distance were selected as the initial screening results, and their corresponding distance values ​​were recorded as a matching degree index. The matching degree index was calculated by normalizing the Mahalanobis distance values ​​and using a negative exponential function to convert the distance values ​​into a matching degree score between 0 and 1; the smaller the distance, the higher the matching degree score. When selecting the initial screening results, an absolute threshold of 5.0 for the Mahalanobis distance was set; fault types exceeding this threshold were not considered to ensure the reliability of the initial screening results.

[0044] 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. 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 to the second smallest distance is less than 0.7, the confidence level of the primary candidate result is considered high; when the ratio is greater than 0.9, the discrimination between the first two candidate results is considered insufficient. Relative position evaluation analyzes the distribution pattern by calculating the area and shape characteristics of the triangle formed by the three candidate results in the feature space. The confidence level is calculated by combining the distance ratio and relative position characteristics using a weighted summation method, where the distance ratio weight is set to 0.6 and the relative position weight is set to 0.4.

[0045] Based on the initial selection results, an adaptive morphological closing operation is performed on the candidate regions to fuse fracture boundaries and fill internal voids. The size of the structuring element in the morphological closing operation is adaptively determined according to the geometric characteristics of the fault region; a 3×1 rectangular structuring element is used for linear fault regions, and a 3×3 circular structuring element is used for planar fault regions. The operation first performs a dilation operation to connect the fractured boundary fragments, and then performs an erosion operation to restore the original size of the region. Through this process, a complete three-dimensional geometric representation of the fault region is generated, including geometric parameters such as region area, perimeter, and minimum bounding rectangle.

[0046] Based on the comprehensive fault type, confidence level, geometric parameters, and spatial location information, a final diagnostic report is automatically generated according to a preset diagnostic report template. The report template includes required fields such as fault type description, confidence level, fault area size, and spatial location coordinates. The fault type description lists three candidate results in descending order of confidence level, each labeled with its respective confidence value. Geometric parameters include the area, maximum size, and shape characteristics of the fault area, with area units converted to square centimeters and size units converted to centimeters. Spatial location coordinates are expressed relative to the transformer reference point, including horizontal distance and vertical height. The diagnostic report also includes a fault severity assessment, automatically classifying faults into three levels—minor, moderate, and severe—based on the size and type of the fault area, and providing corresponding treatment recommendations for each level. The final generated diagnostic report is output in a structured document format, including text descriptions, data tables, and annotated diagrams, ensuring the completeness and readability of the diagnostic results.

[0047] Please see Figure 2 As shown, the image recognition-based transformer fault diagnosis system includes: The dynamic sequence acquisition and quality screening module is used to acquire multimodal image sequences of transformer components in real time, and automatically screen out valid images containing key fault features based on image clarity and component integrity assessment to form an initial image set; The multi-scale template registration module is used to perform multi-scale spatial registration between the initial image set and the pre-constructed standard template of the transformer normal state to generate a reference image with structural alignment. The deformation field inverse derivation module, based on the pixel-level difference between the initial image set and the reference image, inversely derives the displacement vector field characterizing the local deformation of the image; The physical constraint geometric correction module is used to smooth and optimize the displacement vector field under the geometric constraints of the transformer structure, and to perform geometric reconstruction of the initial image based on the optimized displacement field to generate a corrected image of the structure that is true to the structure. The fault diagnosis and report generation module is used to enhance fault features in the gradient domain of the corrected image, extract abnormal region features, identify fault regions based on abnormal region features, and generate the final transformer fault diagnosis report.

[0048] The working principle of this invention is as follows: Accurate fault detection is achieved through multi-stage image processing and feature analysis. Multimodal image sequences of transformer components are simultaneously acquired using a deployed visible light camera and an infrared thermal imager. An initial image set is constructed after quality screening based on sharpness and component integrity. Next, the initial image set is spatially registered with a standard template of the transformer's normal state at multiple scales to generate a reference image, and the displacement vector field is inversely derived based on pixel-level differences. Then, the displacement vector field is smoothed and optimized under the geometric constraints of the transformer structure. Based on the optimized displacement field, the initial image is geometrically reconstructed to obtain a corrected image. Finally, fault features are enhanced in the gradient domain of the corrected image. Fault regions are identified through multi-channel feature extraction and matching with a typical transformer fault feature library. A complete diagnostic report is generated by integrating fault type, confidence level, geometric parameters, and spatial location information. This method achieves accurate identification and quantitative analysis of transformer faults through a complete image processing chain.

[0049] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A transformer fault diagnosis method based on image recognition, characterized in that, Includes the following steps: S1: Real-time acquisition of multimodal image sequences of transformer components, and automatic filtering of valid images containing key fault features based on image clarity and component integrity assessment to form an initial image set; S2: Perform multi-scale spatial registration between the initial image set and the pre-constructed standard template for normal transformer states to generate a reference image with structural alignment; S3: Based on the pixel-level differences between the initial image set and the reference image, the displacement vector field characterizing the local deformation of the image is derived in reverse. S4: Under the geometric constraints 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 corrected image of the structure. S5: Perform fault feature enhancement in the gradient domain of the corrected image, extract abnormal region features, identify fault regions based on abnormal region features, and generate the final transformer fault diagnosis report.

2. The transformer fault diagnosis method based on image recognition according to claim 1, characterized in that, The initial image set specifically includes: By deploying fixed visible light cameras and infrared thermal imagers around the transformer, visible light image sequences and infrared image sequences of transformer components are acquired in a synchronous triggering manner to form a raw multimodal image data stream; For each frame of the original multimodal image data stream, variance calculation based on second-order differential features is used to evaluate image sharpness, and frames that meet the sharpness standard are selected by comparing with a preset dynamic threshold to generate a preliminary set of selected images. Based on the prior three-dimensional geometric information of transformer components, a standard contour projection template for key components is constructed. Each image in the preliminary image set is spatially registered with the template. The overlap ratio of the visible area of ​​the component 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 assessment. By extracting high-frequency gradient features and statistically analyzing their distribution anomalies, images containing potential fault features are initially identified. Finally, these images are collected as valid images to form an initial image set.

3. The transformer fault diagnosis method based on image recognition according to claim 1, characterized in that, The generated reference image with aligned structure specifically includes: Extract key semantic points of transformer component anatomy from each image in the initial image set. Key points include bushing flange edge joints, heat sink fin root connection points, and oil tank weld feature intersections. Construct a Gaussian difference pyramid representation of the standard template for the normal state of a transformer, and establish a scale-invariant spatial coordinate set for the key points of the anatomical semantics at each pyramid level; The anatomical semantic key points are matched bidirectionally with the scale-invariant spatial coordinate set, and the optimal affine transformation parameters are solved by iterative nearest point framework. The optimal affine transformation parameters obtained by the solution are used to perform a global geometric transformation on the initial image set, so that the transformer component structure in the image is precisely aligned with the standard template in space, and the resulting reference image with the aligned structure is output.

4. The transformer fault diagnosis method based on image recognition according to claim 1, characterized in that, The reverse derivation of the displacement vector field characterizing local image deformation specifically includes: A dense pixel correspondence is established between the reference image and the corresponding images in the initial image set. The local similarity of the images is quantified by calculating the normalized cross-correlation value in the neighborhood of each pixel, and an initial similarity distribution map is generated. Anisotropic diffusion filtering is applied to the initial similarity distribution map to enhance the connectivity of high-similarity regions along the image gradient direction, while suppressing local extrema caused by noise, thus forming an optimized similarity distribution map. Based on the optimized similarity distribution map, starting from each pixel position in the reference image, the spatial offset vector of each pixel to the corresponding maximum similarity point is calculated by tracing the maximum similarity path in the initial image. Reassemble the spatial offset vectors of all pixels according to their original positions to construct a complete displacement vector field.

5. The transformer fault diagnosis method based on image recognition according to claim 1, characterized in that, The generated, structurally accurate corrected image specifically includes: Identify regions in the displacement vector field that correspond to the rigid structural components of the transformer, including bushing flanges and tank wall joints, and apply rigid motion constraints to these regions to maintain their geometric invariance; Structural continuity constraints are established for the regions corresponding to non-rigid connections in the displacement vector field. By solving the partial differential equations that satisfy the elastic equilibrium conditions, a smooth displacement field that satisfies the physical structural characteristics of the transformer is obtained. A dense coordinate mapping relationship between the initial image and the reference image is established based on the smoothed displacement field. The initial image is resampled to eliminate geometric distortion while maintaining the integrity of texture features. By verifying the consistency between the geometric dimensions of key transformer components in the reconstructed image and the standard template, a corrected image that conforms to anatomical realism is generated.

6. The transformer fault diagnosis method based on image recognition according to claim 5, characterized in that, The process of obtaining a smooth displacement field that satisfies the physical structural characteristics of the transformer by solving a system of partial differential equations that satisfy the equilibrium conditions of elasticity includes: An anisotropic elastic parameter field for the non-rigid connection region is established based on the material properties of each component of the transformer. In the non-rigidly connected region of the displacement vector field, equilibrium equations based on stress-strain relationship are constructed, and the continuous partial differential equations are discretized into a system of linear equations by the finite difference method. Solve the linear equation system, 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 the displacement field correction amount that satisfies the mechanical equilibrium condition. The displacement field correction is weighted and fused with the original displacement vector field to generate a smooth displacement field that maintains both the geometric deformation characteristics and the structural mechanical continuity of the transformer.

7. The transformer fault diagnosis method based on image recognition according to claim 6, characterized in that, The process involves constructing equilibrium equations based on stress-strain relationships in the non-rigidly connected region of the displacement vector field, and discretizing the continuous partial differential equations into a system of linear equations using the finite difference method. Specifically, this includes: Based on the anisotropic elastic parameter field, a constitutive relationship between the stress tensor and the strain tensor is established at each pixel in the non-rigid connection region, where the strain tensor is determined by the first spatial differential of the displacement vector field. Substituting the stress tensor of each pixel into the equilibrium equation, we obtain a partial differential equation for the second-order spatial differential of the displacement vector field. Discretize the second-order differential terms in the partial differential equation and establish a linear dependency between the displacement vector of each pixel and the displacement vectors of its neighboring pixels. Traverse all pixels in the non-rigidly connected region and assemble the discretized equations into a large sparse linear system of equations containing all unknown displacement vectors in spatial order.

8. The transformer fault diagnosis method based on image recognition according to claim 1, characterized in that, The system identifies fault areas based on abnormal region characteristics and generates a final transformer fault diagnosis report, which includes: In the corrected image enhanced in the gradient domain, multi-channel features are extracted for each candidate anomalous region. The multi-channel features include the hot spot distribution topology, texture spectrum distribution entropy, and boundary tortuosity quantification parameters of the region. The extracted multi-channel features are bidirectionally matched with the typical transformer fault feature library. The initial fault type selection results are determined by calculating the Mahalanobis distance in the feature space, and the confidence level is evaluated. Based on the preliminary selection results, an adaptive morphological closing operation is performed on the candidate region to fuse the fracture boundary and fill the internal holes, generating a complete three-dimensional geometric representation of the fault region. Based on the comprehensive fault type, confidence level, geometric parameters, and spatial location information, the system automatically generates a final diagnostic report containing qualitative conclusions and quantitative analysis data according to a preset diagnostic report template.

9. The transformer fault diagnosis method based on image recognition according to claim 8, characterized in that, The process involves bidirectional matching of the extracted multi-channel features with a typical transformer fault feature library, determining the initial fault type selection by calculating the Mahalanobis distance in the feature space, and evaluating its confidence level. Specifically, this includes: Feature vectors of candidate abnormal regions are constructed based on multi-channel features, and reference fault feature vectors with the same feature dimensions are retrieved from the typical transformer fault feature library. Calculate the Mahalanobis distance between the candidate feature vector and each reference fault feature vector in the feature space. The Mahalanobis distance is obtained based on the covariance relationship between different feature dimensions. The three reference fault types with the smallest Mahalanobis distance were selected as the initial selection results, and the corresponding distance values ​​were recorded as matching degree indicators. 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, forming a set of preliminary selection results for fault types with confidence weights.

10. A transformer fault diagnosis system based on image recognition, characterized in that, The method for performing the image recognition-based transformer fault diagnosis method according to any one of claims 1-9 includes: The dynamic sequence acquisition and quality screening module is used to acquire multimodal image sequences of transformer components in real time, and automatically screen out valid images containing key fault features based on image clarity and component integrity assessment to form an initial image set; The multi-scale template registration module is used to perform multi-scale spatial registration between the initial image set and the pre-constructed standard template of the transformer normal state to generate a reference image with structural alignment. The deformation field inverse derivation module, based on the pixel-level difference between the initial image set and the reference image, inversely derives the displacement vector field characterizing the local deformation of the image; The physical constraint geometric correction module is used to smooth and optimize the displacement vector field under the geometric constraints of the transformer structure, and to perform geometric reconstruction of the initial image based on the optimized displacement field to generate a corrected image of the structure that is true to the structure. The fault diagnosis and report generation module is used to enhance fault features in the gradient domain of the corrected image, extract abnormal region features, identify fault regions based on abnormal region features, and generate the final transformer fault diagnosis report.

Images