Transformer bushing fault location and grading method and system based on infrared image

Abstract

The present application belongs to the technical field of power equipment state monitoring and intelligent diagnosis, and particularly relates to a transformer bushing fault positioning and grading method and system based on infrared images. The method first aligns the structure and segments the axis of the target region of the transformer bushing infrared image according to the axial direction. The segmented images are extracted by the image branch network to obtain the hotspot area feature vector and calculate the temperature profile evidence vector. The profile feature vector is extracted by the profile branch network from the temperature profile evidence vector. The hotspot area feature vector and the corresponding profile feature vector are fused to obtain the fusion feature. The fusion feature of each segment is used to output the fault type and fault probability distribution of each segment to determine the fault segment. The fault feature in the fault segment is extracted, and the fault severity value is calculated based on the fault feature to grade the fault. The present application maps the temperature field to a unified structural coordinate system, suppresses the perspective effect in the infrared inspection scene, and outputs the fault type, location and fault grading.

Description

Technical Field

[0001] This invention belongs to the field of power equipment condition monitoring and intelligent diagnosis technology, specifically relating to a method and system for transformer bushing fault location and classification based on infrared images. Background Technology

[0002] Transformer bushings, as critical insulating components that connect to external leads and withstand high voltage, may experience abnormal heating during long-term operation due to factors such as oil shortage caused by seal aging, creepage or partial discharge caused by moisture and contamination, and increased contact resistance due to poor contact in conductive circuits. In severe cases, this can lead to breakdown or accidents. Currently, transformer bushing maintenance commonly uses online inspection with infrared thermography. However, traditional infrared diagnostic methods rely heavily on temperature rise thresholds, relative temperature differences, or manual experience for interpretation, which suffers from strong subjectivity, sensitivity to changes in shooting distance and angle, and difficulty in outputting precise defect locations and quantifying severity.

[0003] Deep learning-based infrared image recognition can automatically learn features and reduce reliance on manual intervention, but it still faces challenges in engineering environments: temperature distribution differences for the same defect under different viewing angles or radiation reflection conditions lead to insufficient generalization; and outputting only a coarse fault category is insufficient to meet the operational needs for fault location and severity results. There is an urgent need for a transformer bushing infrared diagnostic method that can suppress the influence of viewing angle in infrared inspection scenarios, map the temperature field to a unified structural coordinate system, simultaneously output fault type, location, and severity classification, and possess open-set alarm capabilities. Summary of the Invention

[0004] The purpose of this invention is to address the aforementioned problems in the existing technology by providing a method and system for transformer bushing fault location and classification based on infrared images, which can suppress the influence of viewing angle in infrared inspection scenarios, map the temperature field to a unified structural coordinate system, and simultaneously output the fault type, location, and severity classification.

[0005] To achieve the above objectives, the technical solution of the present invention is as follows:

[0006] In a first aspect, the present invention provides a method for locating and classifying transformer bushing faults based on infrared images, the method comprising:

[0007] First, infrared images of the transformer bushing are acquired and bushing target detection is performed to obtain the bushing target area. Then, key point detection is performed within the bushing target area to obtain the bushing axis direction. Finally, the bushing target area is structurally aligned according to the bushing axis direction to obtain the structural alignment temperature map.

[0008] The structure-aligned temperature map is first expanded radially and then segmented axially to obtain a segmented result, which includes multiple segmented maps. Each segmented map is input into an image branch network to extract image features from the hotspot regions in the segmented map, obtaining the hotspot region feature vector corresponding to each segment. At the same time, the temperature profile evidence vector corresponding to each segmented map is calculated, and the temperature profile evidence vector corresponding to each segmented map is input into a profile branch network to extract profile features, obtaining a profile feature vector.

[0009] First, the hot spot feature vectors of each segment and the corresponding profile feature vectors are gated and fused to obtain the fused features of each segment. Then, the fused features of each segment are input into the pre-built fault identification model to output the fault type and fault probability distribution of each segment. Finally, the fault segment is determined based on the fault type and fault probability distribution of each segment.

[0010] First, the fault features within the fault segment are extracted, and the fault severity value is calculated by weighted summation of the fault features. Then, the severity level is obtained by classifying the fault severity value.

[0011] The hotspot region feature vector includes the hotspot region area. Width ratio of hotspot area The hotspot region's temperature intensity characteristics and texture statistical characteristics; the area of ​​the hotspot region. The expression is:

[0012] ;

[0013] In the above formula, Indicates hotspot areas Mid-pixel coordinates;

[0014] The width ratio of the hotspot area The expression is:

[0015] ;

[0016] In the above formula, , These represent the length of the long side and the length of the short side of the minimum bounding rectangle of the hotspot region, respectively.

[0017] The temperature intensity characteristics of the hotspot region include the maximum temperature rise. Average temperature rise and temperature rise variance The maximum temperature rise The calculation formula is:

[0018] ;

[0019] In the above formula, For pixels Compared to hot spots The average temperature rise;

[0020] The average temperature rise The calculation formula is:

[0021] ;

[0022] In the above formula, This represents the number of pixels within the hotspot area.

[0023] The temperature rise variance The calculation formula is:

[0024] ;

[0025] The texture statistical features of the hotspot areas include the mean temperature gradient. gradient variance The average temperature gradient The calculation formula is:

[0026] ;

[0027] ;

[0028] ;

[0029] ;

[0030] In the above formula, Represents pixels Temperature gradient; , Each pixel Temperature gradients in the horizontal and vertical directions; For pixels Temperature value; For pixels Temperature value; For pixels Temperature value; For pixels Temperature value; This represents the average temperature gradient of the hotspot region;

[0031] The gradient variance The calculation formula is:

[0032] .

[0033] The temperature profile evidence vector includes a segmented average temperature sequence, a segmented maximum temperature sequence, a relative temperature rise sequence, and an axial gradient sequence; the profile feature vector includes the overall trend slope, the low-frequency trend mean, the abrupt change intensity, the relative temperature rise peak value, and the relative temperature rise variance.

[0034] After acquiring the infrared image of the transformer bushing, radiometric correction is first performed to obtain a corrected temperature field map, and then bushing target detection is performed on the corrected temperature field map; the radiometric correction includes:

[0035] First, the axial coordinates of each pixel in the infrared image in the structure alignment coordinate system are calculated, and a reference region is selected based on the axial coordinates. The selection step of the reference region includes: calculating the effective length of the sleeve based on the top and bottom endpoints of the sleeve and constructing an axial unit vector; for any pixel within the target area of ​​the sleeve... The axial coordinates of the sleeve in the structural alignment coordinate system are obtained by axial unit vector projection; wherein, the effective length is the Euclidean distance between the top and bottom endpoints of the sleeve in the infrared image, which is used to characterize the actual axial projection length of the sleeve in the infrared image, and the formula for calculating the effective length is:

[0036] ;

[0037] In the above formula, The effective length; Represents the coordinates of the top endpoint; Indicates the coordinates of the bottom endpoint;

[0038] The formula for calculating the axial unit vector is:

[0039] ;

[0040] In the above formula, Represents the axial unit vector; , These are the coordinate vectors of the bottom and top endpoints of the sleeve, respectively.

[0041] The following formula is used to calculate the pixel points. Projection yields its axial coordinates in the structural alignment coordinate system:

[0042] ;

[0043] In the above formula, Represents pixels Axial coordinates along the sleeve axis;

[0044] Select the location in the structural alignment coordinate system. The middle section, which maintains a clearance distance of not less than 10% of the effective length from both ends of the casing, is used as a reference area.

[0045] Then, with the goal of minimizing the difference in average temperature of the reference region between two adjacent infrared images, the objective function is solved within a preset emissivity range to obtain the optimal emissivity. Based on the optimal emissivity, radiometric inversion is performed on the infrared image of the transformer bushing to obtain the corrected temperature field map; the expression of the objective function is:

[0046] ;

[0047] ;

[0048] In the above formula, , These are the maximum and minimum values ​​within the preset emissivity range, respectively. Emissivity represents the emissivity, which is the optimization variable. The objective function is... , They are respectively time, The average temperature of the reference region in the infrared thermal image obtained at that time; Represents the reference area in an infrared image The number of pixels; Indicates reference area The pixel coordinates in the image.

[0049] The fault characteristics include relative temperature rise peak, axial gradient peak, and highest temperature peak.

[0050] Secondly, the present invention provides a transformer bushing fault location and classification system based on infrared images, the transformer bushing fault location and classification system comprising:

[0051] The structure alignment module is used to first acquire infrared images of transformer bushings and perform bushing target detection to obtain the bushing target area. Then, key point detection is performed within the bushing target area to obtain the bushing axis direction. Finally, the structure alignment of the bushing target area is performed according to the bushing axis direction to obtain the structure alignment temperature map.

[0052] The dual-branch feature extraction module is used to first expand the structure-aligned temperature map radially and then segment it axially to obtain segmented results. The segmented results include multiple segmented maps. Each segmented map is input into the image branch network to extract image features from the hot spots in the segmented map, obtaining the hot spot feature vector corresponding to each segment. At the same time, the temperature profile evidence vector corresponding to each segmented map is calculated. The temperature profile evidence vector corresponding to each segmented map is input into the profile branch network to extract profile features, obtaining the profile feature vector.

[0053] The feature fusion and fault location module is used to first perform gated fusion of the hot spot feature vectors and the corresponding profile feature vectors of each segment to obtain the fused features of each segment, then input the fused features of each segment into the pre-built fault identification model, output the fault type and fault probability distribution of each segment, and then determine the fault segment based on the fault type and fault probability distribution of each segment.

[0054] The fault classification module is used to first extract fault features within fault segments and calculate the fault severity value by weighted summation of the fault features, and then classify the fault severity value to obtain the severity level.

[0055] The hotspot region feature vector includes the hotspot region area. Width ratio of hotspot area The hotspot region's temperature intensity characteristics and texture statistical characteristics; the area of ​​the hotspot region. The expression is:

[0056] ;

[0057] In the above formula, Indicates hotspot areas Mid-pixel coordinates;

[0058] The width ratio of the hotspot area The expression is:

[0059] ;

[0060] In the above formula, , These represent the length of the long side and the length of the short side of the minimum bounding rectangle of the hotspot region, respectively.

[0061] The temperature intensity characteristics of the hotspot region include the maximum temperature rise. Average temperature rise and temperature rise variance The maximum temperature rise The calculation formula is:

[0062] ;

[0063] In the above formula, For pixels Compared to hot spots The average temperature rise;

[0064] The average temperature rise The calculation formula is:

[0065] ;

[0066] In the above formula, This represents the number of pixels within the hotspot area.

[0067] The temperature rise variance The calculation formula is:

[0068] ;

[0069] The texture statistical features of the hotspot areas include the mean temperature gradient. gradient variance The average temperature gradient The calculation formula is:

[0070] ;

[0071] ;

[0072] ;

[0073] ;

[0074] In the above formula, Represents pixels Temperature gradient; , Each pixel Temperature gradients in the horizontal and vertical directions; For pixels Temperature value; For pixels Temperature value; For pixels Temperature value; For pixels Temperature value; This represents the average temperature gradient of the hotspot region;

[0075] The gradient variance The calculation formula is:

[0076] .

[0077] The temperature profile evidence vector includes a segmented average temperature sequence, a segmented maximum temperature sequence, a relative temperature rise sequence, and an axial gradient sequence; the profile feature vector includes the overall trend slope, the low-frequency trend mean, the abrupt change intensity, the relative temperature rise peak value, and the relative temperature rise variance.

[0078] After acquiring the infrared image of the transformer bushing, the structure alignment module first performs radiometric correction on it to obtain a corrected temperature field map, and then performs bushing target detection on the corrected temperature field map.

[0079] The structure alignment module is used to perform radiometric correction according to the following steps: First, for each pixel in the infrared image, its axial coordinate in the structure alignment coordinate system is calculated, and a reference region is selected based on the axial coordinate; the selection step of the reference region includes: calculating the effective length of the sleeve based on the top and bottom endpoints of the sleeve and constructing an axial unit vector; for any pixel within the target area of ​​the sleeve... The axial coordinates of the sleeve in the structural alignment coordinate system are obtained by axial unit vector projection; wherein, the effective length is the Euclidean distance between the top and bottom endpoints of the sleeve in the infrared image, which is used to characterize the actual axial projection length of the sleeve in the infrared image, and the formula for calculating the effective length is:

[0080] ;

[0081] In the above formula, The effective length; Represents the coordinates of the top endpoint; Indicates the coordinates of the bottom endpoint;

[0082] The formula for calculating the axial unit vector is:

[0083] ;

[0084] In the above formula, Represents the axial unit vector; , These are the coordinate vectors of the bottom and top endpoints of the sleeve, respectively.

[0085] The following formula is used to calculate the pixel points. Projection yields its axial coordinates in the structural alignment coordinate system:

[0086] ;

[0087] In the above formula, Represents pixels Axial coordinates along the sleeve axis;

[0088] Select the location in the structural alignment coordinate system. The middle section, which maintains a clearance distance of not less than 10% of the effective length from both ends of the casing, is used as a reference area.

[0089] Then, with the goal of minimizing the difference in average temperature of the reference region between two adjacent infrared images, the objective function is solved within a preset emissivity range to obtain the optimal emissivity. Based on the optimal emissivity, radiometric inversion is performed on the infrared image of the transformer bushing to obtain the corrected temperature field map; the expression of the objective function is:

[0090] ;

[0091] ;

[0092] In the above formula, , These are the maximum and minimum values ​​within the preset emissivity range, respectively. Emissivity represents the emissivity, which is the optimization variable. The objective function is... , They are respectively time, The average temperature of the reference region in the infrared thermal image obtained at that time; Represents the reference area in an infrared image The number of pixels; Indicates reference area The pixel coordinates in the image.

[0093] The fault characteristics include relative temperature rise peak, axial gradient peak, and highest temperature peak.

[0094] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0095] 1. The transformer bushing fault location and classification method of the present invention has the following characteristics: First, it constructs a structural coordinate system by aligning the bushing axis and endpoints geometrically, thereby improving the consistent expression capability across perspectives and scales; second, it forms a consistent sequence of evidence through axial segmented temperature profile reconstruction, thereby enhancing interpretability and supporting fault location output; third, it achieves integrated output of fault type, location, and severity through the fusion of image branches and profile branches, thereby improving on-site safety and usability. Attached Figure Description

[0096] Figure 1 This is a flowchart of the transformer bushing fault location and classification method described in this invention.

[0097] Figure 2 This is a structural block diagram of the transformer bushing fault location and classification system described in this invention. Detailed Implementation

[0098] The present invention will now be described in further detail with reference to specific embodiments and accompanying drawings.

[0099] Example 1:

[0100] See Figure 1 A method for transformer bushing fault location and classification based on infrared images, comprising the following steps:

[0101] S1. First, acquire infrared images of the transformer bushing and perform bushing target detection to obtain the bushing target area. Then, perform key point detection within the bushing target area to obtain the bushing axis direction. Finally, perform structural alignment of the bushing target area according to the bushing axis direction to obtain the structural alignment temperature map.

[0102] Specifically, infrared images of the transformer bushing are acquired using an infrared thermal imager. The target area of ​​the bushing refers to the set of pixels corresponding to the main body of the transformer bushing in the infrared image, including the outer insulation and its skirt structure. Bushing target detection can be achieved using target detection networks such as the YOLO series, Faster R-CNN, and Mask R-CNN. The output of the target detection network is an outer rectangular box or a pixel-level mask. Keypoint detection can be achieved using a keypoint regression model based on a convolutional neural network or a keypoint detection model based on heatmap prediction, such as an encoder-decoder network or an HRNet network. The keypoint coordinates are determined by outputting a keypoint probability heatmap and taking the position of the maximum response. The keypoint coordinates include the positions of the two ends of the bushing, and the straight line connecting the two ends is the axis of the target area of ​​the bushing. An affine transformation is performed on the target area of ​​the bushing based on the axis direction to adjust the axis direction of the target area of ​​the bushing to coincide with the preset axis direction. At the same time, the effective length of the bushing is scaled to a preset length to achieve consistent structural representation across viewpoints. Finally, a structural alignment temperature map is output.

[0103] Specifically, after acquiring the infrared image of the transformer bushing, the reference area in the infrared image is first radiometrically corrected to obtain a corrected temperature field map, and then the bushing target is detected based on the corrected temperature field map.

[0104] The radiation correction includes:

[0105] First, for the infrared image, a structure-aligned coordinate system is constructed, and a reference region is selected within this system. The effective length of the sleeve is calculated based on its top and bottom endpoints, and an axial unit vector is constructed. For any pixel within the target region of the sleeve... The axial coordinates of the sleeve in the structural alignment coordinate system are obtained by axial unit vector projection; wherein, the effective length is the Euclidean distance between the top and bottom endpoints of the sleeve in the infrared image, which is used to characterize the actual axial projection length of the sleeve in the infrared image, and the formula for calculating the effective length is:

[0106] ;

[0107] In the above formula, The effective length; Represents the coordinates of the top endpoint; Indicates the coordinates of the bottom endpoint;

[0108] The formula for calculating the axial unit vector is:

[0109] ;

[0110] In the above formula, Represents the axial unit vector; , These represent the coordinate vectors of the bottom and top ends of the sleeve, respectively.

[0111] The following formula is used to calculate the pixel points. Projection yields its axial coordinates in the structural alignment coordinate system:

[0112] ;

[0113] In the above formula, Represents pixels Axial coordinates along the axial direction of the sleeve. ;

[0114] Select the location in the structural alignment coordinate system. The middle section, which maintains a clearance distance of not less than 10% of the effective length from both ends of the casing, is used as a reference area.

[0115] Then, with the objective of minimizing the difference in average temperature of the reference region between two adjacent infrared images, the optimal emissivity is obtained by solving the objective function within a preset emissivity range using a one-dimensional search or gradient optimization method. Based on the optimal emissivity, the infrared image of the transformer bushing is used for radiometric inversion to obtain a corrected temperature field map; the expression of the objective function is:

[0116] ;

[0117] ;

[0118] In the above formula, , These are the maximum and minimum values ​​within the preset emissivity range, respectively. Emissivity represents the emissivity, which is the optimization variable. The objective function is... , They are respectively time, The average temperature of the reference region in the infrared thermal image obtained at that time; Represents the reference area in an infrared image The number of pixels; Indicates reference area The pixel coordinates in the image.

[0119] S2. The structure-aligned temperature map is first expanded radially and then segmented axially to obtain segmented results. The segmented results include multiple segmented maps. Each segmented map is input into an image branch network to extract image features from the hotspot regions in the segmented map, obtaining the hotspot region feature vector corresponding to each segment. At the same time, the temperature profile evidence vector corresponding to each segmented map is calculated. The temperature profile evidence vector corresponding to each segmented map is input into a profile branch network to extract profile features, obtaining a profile feature vector.

[0120] Specifically, the hotspot region refers to the set of pixels whose temperature is significantly higher than the hotspot threshold, used to characterize localized abnormal heating areas. The hotspot region selection process is as follows: first, based on the reference region... average temperature with temperature standard deviation Constructing hotspot thresholds: ,in Reference area The temperature value is satisfied The pixels constitute hotspot candidate regions. Connectivity analysis is performed on the hotspot candidate regions, and connected regions with an area larger than the target area of ​​the sleeve by 0.1-2% are retained as hotspot regions. .

[0121] Specifically, the hotspot region feature vector includes the hotspot region area. Width ratio of hotspot area The hotspot region's temperature intensity characteristics and texture statistical characteristics; the area of ​​the hotspot region. The expression is:

[0122] ;

[0123] In the above formula, Indicates hotspot areas Mid-pixel coordinates;

[0124] The width ratio of the hotspot area The expression is:

[0125] ;

[0126] In the above formula, , These represent the length of the long side and the length of the short side of the minimum bounding rectangle of the hotspot region, respectively.

[0127] The temperature intensity characteristics of the hotspot region include the maximum temperature rise. Average temperature rise and temperature rise variance The maximum temperature rise The calculation formula is:

[0128] ;

[0129] In the above formula, For pixels Compared to hot spots The average temperature rise;

[0130] The average temperature rise The calculation formula is:

[0131] ;

[0132] In the above formula, This represents the number of pixels within the hotspot area.

[0133] The temperature rise variance The calculation formula is:

[0134] ;

[0135] The texture statistical features of the hotspot areas include the mean temperature gradient. gradient variance The average temperature gradient The calculation formula is:

[0136] ;

[0137] ;

[0138] ;

[0139] ;

[0140] In the above formula, Represents pixels Temperature gradient; , Each pixel Temperature gradients in the horizontal and vertical directions; For pixels Temperature value; For pixels Temperature value; For pixels Temperature value; For pixels Temperature value; This represents the average temperature gradient of the hotspot region;

[0141] The gradient variance The calculation formula is:

[0142] ;

[0143] Specifically, the temperature profile evidence vector includes a segmented average temperature sequence, a segmented maximum temperature sequence, a relative temperature rise sequence, and an axial gradient sequence; let the first... The axial temperature sequence of the casing in a certain segment of the frame is as follows: ,in Represented as for the first A segment in a frame, its discrete points along the axis Temperature value at the location, axial discrete point The location indicates the first segment along the axial direction of the casing. A discrete sampling location; The number of discrete sampling points is defined as the piecewise average temperature. Then continuous The segmented average temperature sequence formed within the frame is Define the highest temperature of the segment as Then continuous The segmented average temperature sequence formed within the frame is Define axial discrete points The relative temperature rise is Robust baseline temperature is The robust baseline temperature is taken as the median of the segmented highest temperature sequence, then the continuous... The segmented relative temperature rise sequence formed within the frame is The segmented relative temperature rise sequence is used to characterize the temperature rise amplitude of local hotspots relative to the overall background and to suppress the influence of the overall ambient temperature drift; the axial gradient is defined as... Then continuous The segmented axial gradient sequence formed within the frame is The segmented average temperature sequence, segmented maximum temperature sequence, relative temperature rise sequence, and axial gradient sequence are concatenated to form a temperature profile evidence vector. This serves as the input to the profile branch network. In this embodiment, the number of segments can be set to 16-64 to balance positioning resolution and statistical stability.

[0144] Specifically, the profile feature vector includes: overall trend slope, low-frequency trend mean, abrupt change intensity, relative temperature rise peak value, and relative temperature rise variance;

[0145] The formula for calculating the overall trend slope is: ,in For all axial discrete points The average coordinates For all axial discrete points Temperature average; via overall trend slope Used to characterize the segmented overall heating or cooling trend;

[0146] The calculation steps for the low-frequency trend mean include: first, extracting axial discrete points using the following formula. Low-frequency trend components at the location: , Discrete points along the axis The moving average window length is centered at the data point, and is generally set in the range of 5-11 sampling points. The larger the window, the higher the smoothness. In this embodiment, a window length of 7 is preferred. A shortened window is used for axial discrete points located at the segment boundaries. Then, all axial discrete points are statistically obtained. Low-frequency trend component mean ;

[0147] The calculation steps for the abrupt change location features include: first calculating the axial discrete points. Second difference at: , the second difference value Points exceeding a preset threshold are identified as mutation points, and then the... Axial discrete points in the frame The temperature gradient at point is denoted as The maximum rate of temperature change among all mutation points is taken as the mutation intensity, which is expressed as: .

[0148] The expression for the profile feature vector is: ,in Represents all axial discrete points The relative temperature rise is determined by the relative temperature rise peak value. Represents all axial discrete points The relative temperature rise variance is determined by the relative temperature rise.

[0149] S3. First, the hot spot feature vectors of each segment and the corresponding profile feature vectors are gated and fused to obtain the fused features of each segment. Then, the fused features of each segment are input into the pre-built fault identification model to output the fault type and fault probability distribution of each segment. Finally, the fault segment is determined based on the fault type and fault probability distribution of each segment.

[0150] Specifically, the gating fusion step includes: referring to the hotspot region feature vector With profile feature vector After mapping to the same dimension, the components are concatenated. The concatenated result is then input into a gating function to calculate the fusion weights. Based on these fusion weights, the feature vectors of hotspot regions are analyzed. With profile feature vector The fusion process is performed to obtain fused features; the expression for the fused features is:

[0151] ;

[0152] ;

[0153] In the above formula, Features of fusion; For weighting; The result of splicing; , These represent the feature vectors for hotspot regions, respectively. , profile feature vector The mapping result; Use the Sigmoid activation function; , These are the gate weight matrix and the bias vector, respectively.

[0154] Specifically, the fault identification model is a multi-task deep learning network that fuses features. As input, an intermediate representation is obtained through a shared hidden layer. The intermediate representation is then passed through a fully connected layer and a Softmax function to output a fault type probability distribution. The fault category with the highest probability in the fault type probability distribution is taken as the fault type predicted by the model. Fault types include, but are not limited to, normal, oil shortage, and local overheating.

[0155] Specifically, let's assume that for the first... The probability distribution of fault types for each segment output is as follows: ,in Indicates the number of fault type categories. Indicates the first The segment belongs to the first The probability of the fault type, then the first The anomaly probability of each segment is: ,in The set of fault types; fault segments are determined based on the anomaly probabilities of all segments: When the fault is segmented The probability of an anomaly is not less than the preset fault probability threshold. When the fault is identified, the fault segment is determined. If the fault segmentation is valid, then the current bushing is determined to have no valid fault segmentation or requires manual verification.

[0156] S4. First, extract the fault features within the fault segment and calculate the fault severity value by weighted summation of the fault features. Then, classify the severity level based on the fault severity value.

[0157] Specifically, the fault characteristics include relative temperature rise peak, axial gradient peak, and highest temperature peak; the fault severity value is obtained by weighted summation according to the following formula:

[0158] ;

[0159] In the above formula, Indicates the severity value of the fault; , , These are the corresponding weighted weights. ; , , Fault segmentation The relative temperature rise peak, axial gradient peak, and highest temperature peak. , , , , , Fault segmentation In the The relative temperature rise, axial gradient, and maximum temperature of the frame; This represents the total number of frames.

[0160] Specifically, the severity level classification based on fault severity values ​​includes pre-setting corresponding thresholds for different severity levels and classifying the severity level based on the pre-set thresholds. For example, severity level classification is performed according to the following formula:

[0161] ;

[0162] In the above formula, , , These represent minor abnormalities, general abnormalities, and severe abnormalities, respectively. , This corresponds to the threshold.

[0163] Example 2:

[0164] See Figure 2 A transformer bushing fault location and classification system based on infrared images includes a structure alignment module, a dual-branch feature extraction module, a feature fusion and fault location module, and a fault classification module.

[0165] The structure alignment module is used to first acquire infrared images of transformer bushings and perform bushing target detection to obtain bushing target areas, then perform key point detection within the bushing target areas to obtain the bushing axis direction, and then perform structure alignment of the bushing target areas according to the bushing axis direction to obtain a structure alignment temperature map.

[0166] The dual-branch feature extraction module is used to first expand the structure-aligned temperature map radially and then segment it axially to obtain segmented results. The segmented results include multiple segmented maps. Each segmented map is input into the image branch network to extract image features from the hot spots in the segmented map, obtaining the hot spot feature vector corresponding to each segment. At the same time, the temperature profile evidence vector corresponding to each segmented map is calculated. The temperature profile evidence vector corresponding to each segmented map is input into the profile branch network to extract profile features, obtaining the profile feature vector.

[0167] The feature fusion and fault location module is used to first perform gated fusion of the hot spot region feature vectors and the corresponding profile feature vectors of each segment to obtain the fused features of each segment, then input the fused features of each segment into the pre-built fault identification model, output the fault type and fault probability distribution of each segment, and then determine the fault segment based on the fault type and fault probability distribution of each segment.

[0168] The fault classification module is used to first extract fault features within the fault segment and calculate the fault severity value by weighted summation of the fault features, and then classify the fault severity value to obtain the severity level.

[0169] The hotspot region feature vector includes the hotspot region area. Width ratio of hotspot area The hotspot region's temperature intensity characteristics and texture statistical characteristics; the area of ​​the hotspot region. The expression is:

[0170] ;

[0171] In the above formula, Indicates hotspot areas Mid-pixel coordinates;

[0172] The width ratio of the hotspot area The expression is:

[0173] ;

[0174] In the above formula, , These represent the length of the long side and the length of the short side of the minimum bounding rectangle of the hotspot region, respectively.

[0175] The temperature intensity characteristics of the hotspot region include the maximum temperature rise. Average temperature rise and temperature rise variance The maximum temperature rise The calculation formula is:

[0176] ;

[0177] In the above formula, For pixels Compared to hot spots The average temperature rise;

[0178] The average temperature rise The calculation formula is:

[0179] ;

[0180] In the above formula, This represents the number of pixels within the hotspot area.

[0181] The temperature rise variance The calculation formula is:

[0182] ;

[0183] The texture statistical features of the hotspot areas include the mean temperature gradient. gradient variance The average temperature gradient The calculation formula is:

[0184] ;

[0185] ;

[0186] ;

[0187] ;

[0188] In the above formula, Represents pixels Temperature gradient; , Each pixel Temperature gradients in the horizontal and vertical directions; For pixels Temperature value; For pixels Temperature value; For pixels Temperature value; For pixels Temperature value; This represents the average temperature gradient of the hotspot region;

[0189] The gradient variance The calculation formula is:

[0190] .

[0191] The temperature profile evidence vector includes a segmented average temperature sequence, a segmented maximum temperature sequence, a relative temperature rise sequence, and an axial gradient sequence; the profile feature vector includes the overall trend slope, the low-frequency trend mean, the abrupt change intensity, the relative temperature rise peak value, and the relative temperature rise variance.

[0192] After acquiring the infrared image of the transformer bushing, the structure alignment module first performs radiometric correction on it to obtain a corrected temperature field map, and then performs bushing target detection on the corrected temperature field map.

[0193] The structure alignment module is used to perform radiometric correction according to the following steps: First, for each pixel in the infrared image, its axial coordinate in the structure alignment coordinate system is calculated, and a reference region is selected based on the axial coordinate; the selection step of the reference region includes: calculating the effective length of the sleeve based on the top and bottom endpoints of the sleeve and constructing an axial unit vector; for any pixel within the target area of ​​the sleeve... The axial coordinates of the sleeve in the structural alignment coordinate system are obtained by axial unit vector projection; wherein, the effective length is the Euclidean distance between the top and bottom endpoints of the sleeve in the infrared image, which is used to characterize the actual axial projection length of the sleeve in the infrared image, and the formula for calculating the effective length is:

[0194] ;

[0195] In the above formula, The effective length; Represents the coordinates of the top endpoint; Indicates the coordinates of the bottom endpoint;

[0196] The formula for calculating the axial unit vector is:

[0197] ;

[0198] In the above formula, Represents the axial unit vector; , These are the coordinate vectors of the bottom and top endpoints of the sleeve, respectively.

[0199] The following formula is used to calculate the pixel points. Projection yields its axial coordinates in the structural alignment coordinate system:

[0200] ;

[0201] In the above formula, Represents pixels Axial coordinates along the sleeve axis;

[0202] Select the location in the structural alignment coordinate system. The middle section, which maintains a clearance distance of not less than 10% of the effective length from both ends of the casing, is used as a reference area.

[0203] Then, with the goal of minimizing the difference in average temperature of the reference region between two adjacent infrared images, the objective function is solved within a preset emissivity range to obtain the optimal emissivity. Based on the optimal emissivity, radiometric inversion is performed on the infrared image of the transformer bushing to obtain the corrected temperature field map; the expression of the objective function is:

[0204] ;

[0205] ;

[0206] In the above formula, , These are the maximum and minimum values ​​within the preset emissivity range, respectively. Emissivity represents the emissivity, which is the optimization variable. The objective function is... , They are respectively time, The average temperature of the reference region in the infrared thermal image obtained at that time; Represents the reference area in an infrared image The number of pixels; Indicates reference area The pixel coordinates in the image.

[0207] The fault characteristics include relative temperature rise peak, axial gradient peak, and highest temperature peak.

Claims

1. A method for transformer bushing fault location and classification based on infrared images, characterized in that: The method for locating and classifying transformer bushing faults includes: Infrared images of transformer bushings are acquired and bushing target detection is performed to obtain the bushing target area. Then, key point detection is performed within the bushing target area to obtain the bushing axis direction. Finally, the bushing target area is structurally aligned according to the bushing axis direction to obtain the structural alignment temperature map. The structure-aligned temperature map is first expanded radially and then segmented axially to obtain a segmented result, which includes multiple segmented maps. Each segmented map is then input into an image branch network to extract image features from the hotspot regions in the segmented map, obtaining a hotspot region feature vector corresponding to each segment. At the same time, the temperature profile evidence vector corresponding to each segmented map is calculated, and the temperature profile evidence vector corresponding to each segmented map is input into a profile branch network to extract profile features, obtaining a profile feature vector. The hotspot region feature vectors of each segment are gated and fused with the corresponding profile feature vectors to obtain the fused features of each segment. The fused features of each segment are then input into the pre-built fault identification model to output the fault type and fault probability distribution of each segment. Finally, the fault segment is determined based on the fault type and fault probability distribution of each segment. First, the fault features within the fault segment are extracted, and the fault severity value is calculated by weighted summation of the fault features. Then, the severity level is obtained by classifying the fault severity value.

2. The method for transformer bushing fault location and classification based on infrared images according to claim 1, characterized in that: The hotspot region feature vector includes the hotspot region area. Width ratio of hotspot area The hotspot region's temperature intensity characteristics and texture statistical characteristics; the area of ​​the hotspot region. The expression is: ； In the above formula, Indicates hotspot areas Mid-pixel coordinates; The width ratio of the hotspot area The expression is: ； In the above formula, , These are the lengths of the longer and shorter sides of the minimum bounding rectangle of the hotspot region, respectively. The temperature intensity characteristics of the hotspot area include the maximum temperature rise. Average temperature rise and temperature rise variance The maximum temperature rise The calculation formula is: ； In the above formula, For pixels Compared to hot spots The average temperature rise; The average temperature rise The calculation formula is: ； In the above formula, This represents the number of pixels within the hotspot area. The temperature rise variance The calculation formula is: ； The texture statistical features of the hotspot areas include the mean temperature gradient. gradient variance The average temperature gradient The calculation formula is: ； ； ； ； In the above formula, Represents pixels Temperature gradient; , Each pixel Temperature gradients in the horizontal and vertical directions; For pixels Temperature value; For pixels Temperature value; For pixels Temperature value; For pixels Temperature value; This represents the average temperature gradient of the hotspot region; The gradient variance The calculation formula is: 。 3. The method for transformer bushing fault location and classification based on infrared images according to claim 1 or 2, characterized in that: The temperature profile evidence vector includes a segmented average temperature sequence, a segmented maximum temperature sequence, a relative temperature rise sequence, and an axial gradient sequence; the profile feature vector includes the overall trend slope, the low-frequency trend mean, the abrupt change intensity, the relative temperature rise peak value, and the relative temperature rise variance.

4. The method for transformer bushing fault location and classification based on infrared images according to claim 1 or 2, characterized in that: After acquiring the infrared image of the transformer bushing, radiometric correction is first performed to obtain a corrected temperature field map, and then bushing target detection is performed on the corrected temperature field map; the radiometric correction includes: First, the axial coordinates of each pixel in the infrared image in the structure alignment coordinate system are calculated, and a reference region is selected based on the axial coordinates. The selection step of the reference region includes: calculating the effective length of the sleeve based on the top and bottom endpoints of the sleeve and constructing an axial unit vector; for any pixel within the target area of ​​the sleeve... The axial coordinates of the sleeve in the structural alignment coordinate system are obtained by axial unit vector projection; wherein, the effective length is the Euclidean distance between the top and bottom endpoints of the sleeve in the infrared image, which is used to characterize the actual axial projection length of the sleeve in the infrared image, and the formula for calculating the effective length is: ； In the above formula, The effective length; Indicates the coordinates of the top endpoint; Indicates the coordinates of the bottom endpoint; The formula for calculating the axial unit vector is: ； In the above formula, Represents the axial unit vector; , These are the coordinate vectors of the bottom and top endpoints of the sleeve, respectively. The following formula is used to calculate the pixel points. Projection yields its axial coordinates in the structural alignment coordinate system: ； In the above formula, Represents pixels Axial coordinates along the sleeve axis; Select the location in the structural alignment coordinate system. The middle section, which maintains a clearance distance of not less than 10% of the effective length from both ends of the casing, is used as a reference area. Then, with the goal of minimizing the difference in average temperature of the reference region between two adjacent infrared images, the objective function is solved within a preset emissivity range to obtain the optimal emissivity. Based on the optimal emissivity, radiometric inversion is performed on the infrared image of the transformer bushing to obtain the corrected temperature field map; the expression of the objective function is: ； ； In the above formula, , These are the maximum and minimum values ​​within the preset emissivity range, respectively. Emissivity represents the emissivity, which is an optimization variable. The objective function is... , They are respectively time, The average temperature of the reference region in the infrared thermal image obtained at that time; Represents the reference area in an infrared image The number of pixels; Indicates reference area The pixel coordinates in the image.

5. The method for transformer bushing fault location and classification based on infrared images according to claim 1 or 2, characterized in that: The fault characteristics include relative temperature rise peak, axial gradient peak, and highest temperature peak.

6. A transformer bushing fault location and classification system based on infrared images, characterized in that: The transformer bushing fault location and classification system includes: The structure alignment module is used to first acquire infrared images of transformer bushings and perform bushing target detection to obtain the bushing target area. Then, key point detection is performed within the bushing target area to obtain the bushing axis direction. Finally, the structure alignment of the bushing target area is performed according to the bushing axis direction to obtain the structure alignment temperature map. The dual-branch feature extraction module is used to first expand the structure-aligned temperature map radially and then segment it axially to obtain segmented results. The segmented results include multiple segmented maps. Then, each segmented map is input into the image branch network to extract image features from the hot spots in the segmented map, obtaining the hot spot feature vector corresponding to each segment. At the same time, the temperature profile evidence vector corresponding to each segmented map is calculated. The temperature profile evidence vector corresponding to each segmented map is input into the profile branch network to extract profile features, obtaining the profile feature vector. The feature fusion and fault location module is used to first perform gated fusion of the hot spot feature vectors and the corresponding profile feature vectors of each segment to obtain the fused features of each segment, then input the fused features of each segment into the pre-built fault identification model, output the fault type and fault probability distribution of each segment, and then determine the fault segment based on the fault type and fault probability distribution of each segment. The fault classification module is used to first extract fault features within fault segments and calculate the fault severity value by weighted summation of the fault features, and then classify the fault severity value to obtain the severity level.

7. The transformer bushing fault location and classification system based on infrared images according to claim 6, characterized in that: The hotspot region feature vector includes the hotspot region area. Width ratio of hotspot area The hotspot region's temperature intensity characteristics and texture statistical characteristics; the area of ​​the hotspot region. The expression is: ； In the above formula, Indicates hotspot areas Mid-pixel coordinates; The width ratio of the hotspot area The expression is: ； In the above formula, , These represent the length of the long side and the length of the short side of the minimum bounding rectangle of the hotspot region, respectively. The temperature intensity characteristics of the hotspot area include the maximum temperature rise. Average temperature rise and temperature rise variance The maximum temperature rise The calculation formula is: ； In the above formula, For pixels Compared to hot spots The average temperature rise; The average temperature rise The calculation formula is: ； In the above formula, This represents the number of pixels within the hotspot area. The temperature rise variance The calculation formula is: ； The texture statistical features of the hotspot areas include the mean temperature gradient. gradient variance The average temperature gradient The calculation formula is: ； ； ； ； In the above formula, Represents pixels Temperature gradient; , Each pixel Temperature gradients in the horizontal and vertical directions; For pixels Temperature value; For pixels Temperature value; For pixels Temperature value; For pixels Temperature value; This represents the average temperature gradient of the hotspot region; The gradient variance The calculation formula is: 。 8. The transformer bushing fault location and classification system based on infrared images according to claim 6 or 7, characterized in that: The temperature profile evidence vector includes a segmented average temperature sequence, a segmented maximum temperature sequence, a relative temperature rise sequence, and an axial gradient sequence; the profile feature vector includes the overall trend slope, the low-frequency trend mean, the abrupt change intensity, the relative temperature rise peak value, and the relative temperature rise variance.

9. The transformer bushing fault location and classification system based on infrared images according to claim 6 or 7, characterized in that: After acquiring the infrared image of the transformer bushing, the structure alignment module first performs radiometric correction on it to obtain a corrected temperature field map, and then performs bushing target detection on the corrected temperature field map. The structure alignment module is used to perform radiometric correction according to the following steps: First, calculate the axial coordinates of the pixels in the infrared image in the structure alignment coordinate system and select a reference area based on the axial coordinates. The selection step of the reference region includes: calculating the effective length of the sleeve based on the top and bottom endpoints of the sleeve and constructing an axial unit vector; for any pixel point within the target area of ​​the sleeve... The axial coordinates of the sleeve in the structural alignment coordinate system are obtained by axial unit vector projection; wherein, the effective length is the Euclidean distance between the top and bottom endpoints of the sleeve in the infrared image, which is used to characterize the actual axial projection length of the sleeve in the infrared image, and the formula for calculating the effective length is: ； In the above formula, The effective length; Indicates the coordinates of the top endpoint; Indicates the coordinates of the bottom endpoint; The formula for calculating the axial unit vector is: ； In the above formula, Represents the axial unit vector; , These are the coordinate vectors of the bottom and top endpoints of the sleeve, respectively. The following formula is used to calculate the pixel points. Projection yields its axial coordinates in the structural alignment coordinate system: ； In the above formula, Represents pixels Axial coordinates along the sleeve axis; Select the location in the structural alignment coordinate system. The middle section, which maintains a clearance distance of not less than 10% of the effective length from both ends of the casing, is used as a reference area. Then, with the goal of minimizing the difference in average temperature of the reference region between two adjacent infrared images, the objective function is solved within a preset emissivity range to obtain the optimal emissivity. Based on the optimal emissivity, radiometric inversion is performed on the infrared image of the transformer bushing to obtain the corrected temperature field map; the expression of the objective function is: ； ； In the above formula, , These are the maximum and minimum values ​​within the preset emissivity range, respectively. Emissivity represents the emissivity, which is an optimization variable. The objective function is... , They are respectively time, The average temperature of the reference region in the infrared thermal image obtained at that time; Represents the reference area in an infrared image The number of pixels; Indicates reference area The pixel coordinates in the image.

10. The transformer bushing fault location and classification system based on infrared images according to claim 6 or 7, characterized in that: The fault characteristics include relative temperature rise peak, axial gradient peak, and highest temperature peak.

Images