Insulator grading method, system and device based on electroluminescent images

Abstract

The present application relates to the technical field of insulator identification, in order to solve the problem of low efficiency and inaccuracy of existing methods in classifying the use state of insulators, the present application provides an insulator classification method, system and device based on electroluminescence images, which trains a YOLOv8 model using a set of insulator string electroluminescence images, obtains an insulator detection model, then uses the insulator detection model to detect the to-be-detected image, obtains the boundary box corresponding to the piece position order of each insulator piece, then automatically extracts the effective light-emitting ROI of the insulator piece based on the boundary box in the color space, and uses the effective light-emitting ROI and the voltage of the corresponding insulator piece to obtain a voltage mapping model, and then obtains the voltage estimate value of the insulator piece, and finally outputs the classification result of the insulator piece, which assists in the later maintenance of the insulator.

Description

Technical Field

[0001] This invention relates to the field of insulator identification technology, and in particular to an insulator grading method, system and device based on electroluminescent images. Background Technology

[0002] Insulators are core components in overhead transmission lines, serving both mechanical support and electrical isolation functions. Under the combined effects of long-term mechanical loads, electrical stress, environmental erosion, pollution deposition, thermal cycling, and lightning strikes, insulators may experience material degradation, internal microcracks, localized carbonization, interface debonding, or through-breakdown, leading to a significant decrease in insulation resistance. When an insulator essentially loses its insulating capacity, the voltage across its terminals approaches zero; this is commonly referred to as a zero-value insulator. Zero-value insulators distort the voltage distribution across the entire string of insulators, causing adjacent healthy insulators to experience higher voltage stress, thereby increasing the risk of flashover, string failure, and line tripping.

[0003] Traditional methods for detecting zero-value insulators mainly include spark gap method, infrared thermography, electric field distribution measurement method, ultraviolet imaging method, and manual tower climbing inspection. Spark gap method usually requires inspection of each insulator individually, which is labor-intensive, inefficient, and has high safety risks; infrared thermography relies on the temperature rise caused by defects and is not sensitive enough to early low-value defects and defects with insignificant voltage heating; electric field distribution measurement method requires the deployment of electric field sensors or special probes, combined with numerical inversion analysis, making the system complex and difficult to deploy on a large scale; ultraviolet imaging method mainly reflects corona discharge and is not sensitive to internal zero-value or low-value defects.

[0004] Currently, some existing technologies involve applying an electroluminescent coating to the surface or connection area of ​​the insulator, allowing manual observation of the luminescence intensity and judgment of abnormalities using a simple brightness threshold, or establishing a curve between brightness and voltage / electric field under a calibration environment to determine zero-value insulators. However, these methods have the following problems:

[0005] (1) Zero value judgment based on electroluminescent images relies heavily on the single dimension of luminescence brightness. In actual operating environments, ambient light, nighttime background, camera exposure, gain, white balance, shooting distance, lens angle, coating thickness, coating aging, and surface dirt can all change the absolute brightness value in the image, making it easy for fixed brightness thresholds or simple calibration curves to drift, thus causing misjudgment or missed judgment.

[0006] (2) Simple threshold segmentation is difficult to reliably extract the effective area that truly represents the electroluminescent coating. In the field insulator string image, steel caps, steel feet, umbrella skirts, background sky, wires, hardware and local reflections will all appear in the image at the same time, affecting the accuracy of the effective area extraction of the electroluminescent image and resulting in inaccurate results.

[0007] (3) Existing research is mostly limited to the stage of coating light emission to manual visual judgment, which is labor-intensive, has large errors and low efficiency. It lacks a complete and intelligent analysis link and cannot effectively improve work efficiency.

[0008] Therefore, there is an urgent need to design an insulator classification method, system, and device based on electroluminescent images. Summary of the Invention

[0009] To address the shortcomings of existing technologies, the present invention aims to provide an insulator grading method, system, and device based on electroluminescent images. This method trains a YOLOv8 model using a set of electroluminescent images of insulator strings to obtain an insulator detection model. The insulator detection model is then used to detect the images to be detected, obtaining bounding boxes corresponding to the positional order of each insulator piece. Based on these bounding boxes, effective emitting regions of interest (ROIs) for each insulator piece are automatically extracted using color space clustering. A voltage mapping model is obtained using the effective ROIs and the corresponding voltage of the insulator pieces, yielding a voltage prediction value for the insulator string. Based on the voltage prediction value and a preset anomaly threshold, the effective emitting ROIs are determined to be valid or invalid images. If a valid image is found, the insulator pieces are graded based on the voltage prediction value and the normal average voltage value of the insulator string, and the grading results are output to assist in the subsequent maintenance of the insulator pieces.

[0010] Unlike methods based solely on brightness thresholds, this invention does not directly read the average grayscale or brightness of an image. Instead, it constructs multidimensional color features that include hue, saturation, and brightness distributions, and eliminates background interference through target detection and adaptive segmentation of luminous area before feature extraction.

[0011] To achieve the above objectives, the technical solution adopted by the present invention is as follows: The first objective of this invention is to provide a method for classifying insulators based on electroluminescent images, comprising: The image to be detected is obtained, and the image is detected by the insulator detection model to obtain the bounding box of each insulator piece in the insulator string. The order of the bounding boxes corresponds to the position order of the insulator pieces. The image to be detected is the electroluminescent image of the insulator string. Clustering algorithms are used to cluster the image pixels within the bounding box to obtain a primary luminous image. The primary luminous image is then processed to obtain candidate luminous areas. The largest connected component among the candidate luminous areas is selected as the final luminous image. Based on the area ratio of the final luminous image within the corresponding bounding box, it is determined whether the effective luminous information of the final luminous image is sufficient; If insufficient, the final luminescent image is determined to have low confidence and is marked as requiring re-inspection; If sufficient, the final luminous image is determined to be an effective luminous ROI; a color space mapping transformation is performed on the effective luminous ROI to obtain a multidimensional color feature vector of the effective luminous ROI; the multidimensional color feature vector is used as the model input, and the voltage of the insulator sheet corresponding to the effective luminous ROI is used as the label to train a machine learning regression model to obtain a voltage mapping model. The voltage mapping model outputs the voltage estimate of the insulator piece corresponding to the effective ROI, thereby obtaining the voltage estimate of the insulator string. Based on the voltage estimate of the insulator string and a preset anomaly threshold, the effective ROI is determined to be a valid image or an invalid image. If the image is invalid, it is determined that the image in the bounding box corresponding to the effective ROI is invalid or the voltage mapping model output is abnormal. If it is a valid image, based on the voltage prediction of the insulator piece and the normal average voltage value of the insulator string, the voltage ratio of the insulator piece corresponding to the valid ROI is calculated, and the grading result of each insulator piece in the image to be detected is obtained according to the voltage ratio.

[0012] Preferably, the step of clustering the image pixels within the bounding box using a clustering algorithm to obtain a primary luminous image, and processing the primary luminous image to obtain candidate luminous areas, includes: Clustering algorithms are used to cluster the image pixels within the bounding box to determine multiple cluster centers. The comprehensive chromaticity value of each cluster center is calculated, and the cluster center corresponding to the largest comprehensive chromaticity value is determined as the target cluster center. The primary luminous image is then determined based on the target cluster center. The primary luminescent image is filled with small holes and broken areas, and isolated noise is removed. The outer contour of the primary luminescent image is extracted, and the candidate luminescent area is determined based on the outer contour.

[0013] Preferably, determining whether the effective luminescence information of the final luminescence image is sufficient based on the area ratio of the final luminescence image within the corresponding bounding box includes: Compare the area ratio with the preset threshold; If the area ratio is less than a preset threshold, it is determined that the luminous information of the final luminous image is insufficient; If the area ratio is greater than or equal to a preset threshold, the final luminous image is determined to have sufficient luminous information.

[0014] Preferably, the effective emitting ROI undergoes a color space mapping transformation to obtain a multidimensional color feature vector of the effective emitting ROI, including: The effective emitting ROI is subjected to a color space mapping transformation to map it to the HSV color space; The pixel distribution of the effective emitting ROI in the H channel, S channel and V channel of the HSV color space is statistically analyzed respectively; Based on the statistical pixel distribution results, a multidimensional color feature vector of the effective luminous ROI is established.

[0015] Preferably, the voltage prediction value of the insulator sheet corresponding to the effective emitting ROI is output through the voltage mapping model, thereby obtaining the voltage prediction value of the insulator string, including: The multidimensional color feature vector of the effective emitting ROI is input into the voltage mapping model, and the voltage prediction value of the insulator corresponding to the effective emitting ROI is output. The estimated voltage values ​​of all the insulator discs on the insulator string are added together to obtain the estimated voltage value of the insulator string.

[0016] Preferably, determining whether the effective ROI is a valid or invalid image based on the voltage prediction of the insulator string and a preset anomaly threshold includes: If the voltage estimate is less than or equal to a preset abnormal threshold, the effective ROI is determined to be an invalid image; If the voltage estimate is greater than the preset abnormal threshold, the effective ROI is determined to be a valid image.

[0017] Preferably, based on the voltage prediction of the insulator disc and the normal equal voltage value of the insulator string, the voltage ratio of the insulator disc corresponding to the effective ROI is calculated, including: The normal equal voltage value of the insulator string is determined by the ratio of the total voltage of the insulator string to the number of insulator discs in the insulator string. Calculate the normalized voltage U of the insulator disc after the total voltage constraint of the insulator string. i The calculation formula is: U i =U total ·û i / ( +ε); where û i Let û be the voltage prediction value of the i-th insulator piece output by the voltage mapping model. j Let U be the voltage prediction value of the j-th insulator piece output by the voltage mapping model, n be the number of insulator pieces in the insulator string, ε be a small constant, and U be the voltage prediction value of the j-th insulator piece. total Let be the total voltage across the insulator string, and j be the summation subscript; The voltage ratio of the insulator disc is obtained by comparing the normalized voltage with the normal average voltage.

[0018] Preferably, obtaining the grading result of each insulator disc in the image to be detected based on the voltage ratio includes: The voltage ratio is compared with the first judgment threshold and the second judgment threshold; If the voltage ratio is less than the first judgment threshold, the insulator is determined to be zero or severely degraded. If the first judgment threshold is less than or equal to the voltage ratio and less than the second judgment threshold, the insulator is determined to be a low-value deteriorated insulator. If the voltage ratio is greater than or equal to the second judgment threshold, the insulator is determined to be a normal insulator.

[0019] The second objective of this invention is to provide an insulator grading system based on electroluminescent images, wherein the system is implemented based on an insulator grading method using electroluminescent images, and the system includes: The image detection module is used to acquire the image to be detected. It detects the image to be detected through the insulator detection model to obtain the bounding box of each insulator piece in the insulator string. The order of the bounding boxes corresponds to the position order of the insulator pieces. The image to be detected is the electroluminescent image of the insulator string. The first judgment module is used to cluster the image pixels within the bounding box using a clustering algorithm to obtain a primary luminous image, process the primary luminous image to obtain candidate luminous areas, and select the largest connected component among the candidate luminous areas as the final luminous image. Based on the area ratio of the final luminous image within the corresponding bounding box, it is determined whether the effective luminous information of the final luminous image is sufficient; If insufficient, the final luminescent image is determined to have low confidence and is marked as requiring re-inspection; If sufficient, the final luminous image is determined to be an effective luminous ROI; a color space mapping transformation is performed on the effective luminous ROI to obtain a multidimensional color feature vector of the effective luminous ROI; the multidimensional color feature vector is used as the model input, and the voltage of the insulator sheet corresponding to the effective luminous ROI is used as the label to train a machine learning regression model to obtain a voltage mapping model. The output judgment module is used to output the voltage prediction value of the insulator piece corresponding to the effective emitting ROI through the voltage mapping model, thereby obtaining the voltage prediction value of the insulator string, and judging whether the effective emitting ROI is a valid image or an invalid image based on the voltage prediction value of the insulator string and a preset abnormal threshold. If the image is invalid, it is determined that the image in the bounding box corresponding to the effective ROI is invalid or the voltage mapping model output is abnormal. If it is a valid image, based on the voltage prediction of the insulator piece and the normal average voltage value of the insulator string, the voltage ratio of the insulator piece corresponding to the valid ROI is calculated, and the grading result of each insulator piece in the image to be detected is obtained according to the voltage ratio.

[0020] A third objective of the present invention is to provide an electronic device, the device comprising: At least one processor, and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor to enable the at least one processor to perform an insulator grading method based on electroluminescent images.

[0021] The beneficial effects of this invention are: This invention discloses an insulator grading method based on electroluminescent images. Compared with the prior art, the improvement of this invention lies in: (1) The method of this invention, through a complete link of image acquisition, target detection, luminous area extraction, color feature, voltage normalization, and state diagnosis, elevates the detection of insulator pieces from visualization to a quantifiable, verifiable, and reportable intelligent grading method. Its grading results can distinguish between single-piece luminous abnormalities and insulators with abnormal distribution across the entire string. It is suitable for engineering scenarios in actual high-voltage lines where insulators exhibit inter-piece coupling, electric field distortion, and series-end effects. This solves the problem that existing insulator piece electroluminescence detection mainly relies on manual visual inspection and simple testing, resulting in low detection accuracy and high difficulty. Moreover, it does not require the installation of electric field sensors, power supply cables, and communication modules on each insulator piece. It only utilizes existing visible light imaging equipment to acquire insulator electroluminescence images and completes the diagnosis directly at the ground end or edge end, making it highly adaptable to drone, robot, and manual inspection scenarios.

[0022] (2) In view of the problem that the single brightness threshold is easy to drift, the present invention uses the hue, saturation and brightness of the light-emitting area of ​​the insulator as features and performs histogram normalization to avoid the problem of low detection accuracy caused by relying solely on absolute brightness value and failing to identify interference information. At the same time, the present invention also uses the total voltage of the entire string of insulators to normalize the individual insulator pieces, further reducing the scale deviation caused by the overall exposure change and ensuring the accuracy of the output results.

[0023] (3) In view of the problem that the luminescent area of ​​the electroluminescent image is difficult to be stably extracted, the present invention first uses a bounding box to limit the analysis range of the insulator sheet, and then uses the cluster center corresponding to the largest comprehensive chromaticity value in the red-green-yellow-blue color space as the determination of the coating luminescent area. This operation reduces the interference of the sky, hardware, umbrella skirt and reflective background in the electroluminescent image on the extraction and determination of the coating luminescent area. Attached Figure Description

[0024] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. In the accompanying drawings: Figure 1 This is a flowchart of the insulator grading method provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the boundary frame of five single insulators on an insulator string provided in an embodiment of the present invention; Figure 3 This is the output interface of the insulator grading system structure provided in the embodiments of the present invention. Detailed Implementation

[0025] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. The preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0026] Example: like Figures 1 to 3 As shown, this embodiment of the invention provides an insulator grading method based on electroluminescent images, the method comprising: S1. Obtain the image to be detected, and use the insulator detection model to detect the image to be detected, and output the bounding box corresponding to the position sequence of each insulator piece in the insulator string; wherein, the image to be detected is the electroluminescent image of the insulator string.

[0027] Specifically, an electroluminescent image set of insulator strings is established, and an image to be detected is acquired. The depth detection model is trained using the electroluminescent image set of insulator strings to obtain an insulator detection model. The insulator detection model is then used to detect the image to be detected, and the bounding boxes corresponding to the position sequence of each insulator piece in the insulator string are obtained. In this embodiment, the image to be detected is an electroluminescent image of an insulator string.

[0028] In this embodiment, the process of establishing an electroluminescence image set for insulator strings is as follows: First, an electroluminescent composite coating is applied to a single insulator sheet or its connection area on the insulator string to be tested. This coating may include ZnS:Cu luminescent powder (copper-doped zinc sulfide phosphor), an insulating resin matrix, and necessary dispersing agents. The coating location may be between the steel cap and the skirt of the insulator sheet, the area near the steel foot, the area near the connecting hardware, or other locations that can reflect the potential difference or local electric field of the insulator sheet. Multiple electroluminescence images of the insulator string can be acquired using a common visible light camera, a low-light camera, an industrial camera, a multispectral camera, or a dedicated imaging device with filters, and are divided into training images and images to be tested. Among them, the training images are electroluminescence images of the insulator string acquired under known single-sheet voltage, whole string voltage, or equivalent electric field conditions in the field, while the images to be tested are electroluminescence images of the insulator string acquired under unknown single-sheet voltage, known whole string voltage, or equivalent electric field conditions in the same field.

[0029] Training images are named using the format "voltage value + kV", such as 0kV, 3kV, 6kV, 9kV, 12kV, 15kV, 18kV, 21kV, 24kV, etc., to establish supervised learning samples between color features and voltage. During the sample construction phase, metadata such as exposure time, gain, white balance mode, shooting distance, shooting angle, ambient illuminance, insulator disc type, coating batch, and surface contamination level of the training images can also be recorded. This metadata can be used for subsequent error analysis of insulator disc grading results, and can also serve as optional extended inputs to further train more complex mapping models or for image quality screening.

[0030] Then, the LabelMe annotation tool was used to annotate the connection regions of each insulator sheet or its coating in the training images, and the annotated data was converted into a target detection training format, such as a txt file. The converted annotated data was divided into a training set and a validation set, with 80% for training and 20% for validation. The training set was used to train the deep detection model, allowing the model to continuously adjust its parameters. The validation set was used to evaluate the model's performance in real time during training, guiding the adjustment of model parameters and avoiding overfitting, ultimately resulting in the insulator detection model. Preferably, the depth detection model can be YOLOv8 (version 8 single-stage real-time object detection algorithm), YOLOv10 (version 10 single-stage real-time object detection algorithm), Faster R-CNN (faster region convolutional neural network), SSD (single-stage multi-box detector), RetinaNet (retina network), DETR (detection transformer), RT-DETR (real-time detection transformer), Mask R-CNN (masked region convolutional neural network), instance segmentation network, and semantic segmentation network. If the pose of the insulator string is fixed on site, template matching, traditional contour detection, and manual initial bounding box combined with tracking algorithms can also be used. In this embodiment, the depth detection model is YOLOv8.

[0031] Next, the image to be detected is input into the insulator detection model for detection. The model outputs the primary bounding box of each insulator piece in the insulator string and the detection confidence score of the primary bounding box. Detect confidence and detection confidence threshold By comparing the results, we can obtain the final trustworthy bounding boxes. .

[0032] In this embodiment, the detection confidence threshold is... Set to 0.5, when the detection confidence of the i-th primary bounding box is... < When the bounding box is considered low-quality, it is discarded or marked for re-inspection; when the detection confidence of the i-th primary bounding box is... ≥ When this happens, the bounding box is considered a trustworthy bounding box, such as... Figure 2 As shown, the bounding box is preserved and proceeds to subsequent steps. The detection confidence threshold is determined under different shooting distances, image sharpness, insulator type, and training sample size. It can be adjusted within the range of 0.3-0.7; preferably, in scenarios where there is greater sensitivity to missed detections, the detection confidence threshold is [value missing]. A value of 0.3-0.4 is suitable for scenarios that are more sensitive to false detections, representing the detection confidence threshold. A value of 0.6-0.7 is acceptable.

[0033] The expression for the bounding box is: =( , , , ),in, Let x be the x-coordinate of the top-left pixel of the i-th bounding box. Let be the ordinate of the top-left pixel of the i-th bounding box. Let x be the x-coordinate of the bottom right pixel of the i-th bounding box. Let be the ordinate of the bottom right corner pixel of the i-th bounding box. The center point coordinates of the corresponding insulator piece are obtained through this bounding box. The specific method is as follows: =((x 1i +x 2i ) / 2,(y 1i +y 2i ) / 2).

[0034] Finally, based on the center point coordinates of each insulator disc Next, determine the positional order of each insulator disc on the insulator string. Specifically, if the insulator string in the image to be inspected is distributed longitudinally, then follow the sequence... Sort by ascending order of the ordinate; if the insulator strings in the image to be detected are distributed horizontally, then sort by... Sort by x-coordinate in ascending order, based on center point coordinates. This allows us to obtain the position sequence of the first to nth insulator discs on the insulator string, where n is the number of insulator discs in the insulator string. Finally, we obtain the bounding box corresponding to the position sequence of each insulator disc in the insulator string. That is, each bounding box has a sequence number, and this sequence number is consistent with the position sequence of the insulator discs corresponding to the bounding box.

[0035] In another embodiment, the insulator detection model not only outputs the bounding boxes of the insulator pieces, but also simultaneously outputs the connection hardware area, skirt transition area, and steel cap position of each insulator piece as auxiliary target categories to correct the bounding box of each insulator piece. and center point coordinates i This way, even if the entire string of insulators is obscured by a complex background, the position of the insulator segments can still be inferred from the local structure.

[0036] S2. Cluster the image pixels within the bounding box using a clustering algorithm to obtain a primary luminous image. Process the primary luminous image to obtain candidate luminous areas, and select the largest connected component among the candidate luminous areas as the final luminous image. This includes the following steps: S201. Use a clustering algorithm to cluster the image pixels within the bounding box, determine multiple cluster centers, calculate the comprehensive chromaticity value of each cluster center, determine the cluster center corresponding to the largest comprehensive chromaticity value as the target cluster center, and determine the primary luminous image based on the target cluster center.

[0037] Specifically, let the set of image pixels within the bounding box be denoted as , ={ ... },in, Let i be the first pixel within the i-th bounding box. Let N be the Nth pixel within the i-th bounding box, where N is the total number of pixels within the i-th bounding box; let K be the number of cluster centers in the clustering algorithm, for the image pixel set... Perform the K-means clustering algorithm to obtain K cluster centers. ,and =( , , ),in, For the luminance component, For red and green components, The yellow-blue component. Based on and Calculate the comprehensive chromaticity value of the kth cluster center. The calculation formula is: , The distance of the k-th cluster center relative to the neutral color point on the Lab chromaticity plane is represented by the comprehensive chromaticity value. The cluster center corresponding to the largest comprehensive chromaticity value is determined as the target cluster center. Pixels belonging to the cluster corresponding to the target cluster center are marked as foreground, and the remaining pixels are marked as background, resulting in a binary mask of the luminous region within the bounding box. The mask is then mapped back to the original image size according to the bounding box coordinates, and the region outside the bounding box is uniformly set as background, resulting in the primary luminous image. Since the Lab color space (luminance component L, red-green component a, yellow-blue component b, CIELAB) color encoding (Lab encoding) of the open-source computer vision library (OpenCV) uses 128 to represent neutral colors, the comprehensive chromaticity value calculation can more accurately characterize the degree of chromaticity deviation of the luminous region relative to the background.

[0038] In another embodiment, one or a combination of at least two of the following methods may be used to replace the technique of determining the primary luminescence candidate region through clustering algorithms: HSV color space thresholding, luminance-red-green-yellow-blue color space (Lab) thresholding, Gaussian mixture model (GMM) clustering, mean-shift clustering, iterative graph cut (GrabCut), superpixel segmentation, color index segmentation, U-Net segmentation network, and attention heatmap-based region extraction. This substitution will not affect the subsequent steps of this application.

[0039] S202, fill in the small holes and broken areas of the primary luminescent image and remove isolated noise, extract the outer contour of the primary luminescent image, and determine the candidate luminescent area based on the outer contour.

[0040] Specifically, morphological opening operations are performed on the primary luminescent image to remove isolated noise, closing operations are performed to fill small holes and broken areas within the primary luminescent image, and the outer contour of the primary luminescent image is extracted. Based on this outer contour, candidate luminescent areas are finally obtained, and the largest connected region among the candidate luminescent areas is selected as the final luminescent image.

[0041] S3. Based on the area ratio of the final luminous image within the corresponding bounding box, determine whether the effective luminous information of the final luminous image is sufficient. If insufficient, determine that the final luminous image has low confidence and mark it as needing re-examination. If sufficient, determine that the final luminous image is a valid luminous ROI and proceed to step S4. Specifically, this includes the following steps: Calculate the area ratio of the final emission image within the corresponding bounding box. The area ratio is then compared with a preset threshold to determine whether the effective luminous information of the final luminous image is sufficient.

[0042] Specifically, the formula for calculating the area ρi of the final luminescent image within the bounding box is: = / ,in, The area of ​​the final luminescent image. This represents the area of ​​the bounding box. If... Less than the preset threshold If the final emission image is deemed to have insufficient emission information, the corresponding insulator sheet is considered to have insufficient effective emission information, and the electroluminescent image of the insulator sheet is output as low confidence or requires re-inspection. Greater than or equal to the preset threshold If the final emission image is deemed to have sufficient emission information, it is considered that the corresponding insulator sheet has sufficient effective emission information and requires further detection. The corresponding final emission image is then considered the effective emission ROI. This step eliminates the need for manual trimming of the emission area in the electroluminescent image piece by piece, and can stably obtain the effective emission area in the electroluminescent image used for voltage / electric field inversion even in complex backgrounds.

[0043] In this embodiment of the application, a preset threshold is used. A score of 0.02 is used, meaning that when the area of ​​the final luminescent image is less than 2% of the corresponding bounding box area, the final luminescent image is considered to have insufficient effective luminescence information and is marked as low confidence, pending re-examination; under different camera resolutions, coating widths, and shooting distances, It can be adjusted within the range of 0.01 to 0.05.

[0044] S4. Perform color space mapping transformation on the effective emitting ROI to obtain the multidimensional color feature vector of the effective emitting ROI. Use the multidimensional color feature vector as the model input and the voltage of the insulator corresponding to the effective emitting ROI as the label to train the machine learning regression model and obtain the voltage mapping model.

[0045] In this embodiment of the application, the idea of ​​performing color space mapping transformation on the effective emitting ROI to obtain the multidimensional color feature vector of the effective emitting ROI is as follows: perform HSV color space mapping transformation on the effective emitting ROI to map it to the HSV color space, and respectively count the pixel distribution of the effective emitting ROI in the H channel, S channel and V channel of the HSV color space, and establish the multidimensional color feature vector of the effective emitting ROI based on the statistical pixel distribution results.

[0046] Specifically, the effective emitting ROI undergoes HSV color space mapping transformation to obtain the H, S, and V channels of the effective emitting ROI in the HSV color space. The pixel distribution of each channel is then statistically analyzed to obtain the corresponding normalized histograms. Taking the H channel as an example, let its bin number be... m=1,2,3,... The pixel count of the m-th bin is Then the normalized histogram of the m-th bin is: = / ( +ε), where, Let be the normalized value of the m-th bin, t be the traversal index of the summation operation used to sequentially refer to each bin in the histogram, valid only within the calculation range of the summation symbol, and ε be a small constant to prevent the denominator from being zero. The normalized values ​​of the S and V channels are calculated in the same way as the H channel, thus obtaining the normalized value of the S channel. and the normalized value of the V channel This application will not elaborate further on this point. Preferably, =30, =10, =10, where, This represents the number of bins in the S-channel. This represents the number of bins for the V channel. The number of bins in the histogram is not limited to 30 / 10 / 10 and can be adjusted according to the calibration sample size and imaging resolution. The normalized value... , as well as Constructing a multidimensional color feature vector , =[ ,..., , ,..., , ,..., This yields a multi-dimensional color feature vector of the effective luminous ROI, which is 50-dimensional; where, This is the normalized value of the first bin in the H channel. The normalized value for the thirtieth bin. This is the normalized value of the first bin in the S channel. This is the normalized value of the tenth bin in the S channel. This is the normalized value of the first bin of the V channel. The normalized values ​​for the ten bins of the V channel.

[0047] The technical significance of this step lies in the fact that the H channel reflects the main band or hue shift of the emitted color in the electroluminescent image, the S channel reflects the purity of the emitted light and the degree of light / background blending in the electroluminescent image, and the V channel reflects the brightness distribution in the electroluminescent image; the three channels together describe the state of the electroluminescent image. Through the effective emission ROI processed by the distribution characteristics and normalization, even if the overall brightness amplitude changes due to camera exposure or shooting distance, its color distribution pattern can still provide more stable information than a single brightness mean.

[0048] In another embodiment, the present invention may further introduce at least one of color moments, local texture statistics, color correlation maps, and depth convolution features on the basis of the normalized histograms of the H, S, and V channels, and perform concatenation or weighted fusion with the normalized histograms of the three channels (H, S, and V). This can maintain the interpretability of the method of the present invention and improve the fitting ability to complex nonlinear electroluminescence variation laws when the number of samples is sufficient.

[0049] In another embodiment, one or at least two combinations of red-green-blue (RGB), luminance, red-green, yellow-blue color space (Lab), and joint features of multiple color spaces can be used to replace the HSV color space mapping transformation to obtain a multidimensional color feature vector of an effective luminous ROI. The multidimensional color feature vector obtained after the replacement will also be different. Then, a machine learning regression model is trained based on the replaced multidimensional color feature vector.

[0050] In this embodiment, the process of combining the multidimensional color feature vector of the effective emitting ROI and the voltage of the insulator corresponding to the effective emitting ROI to train a machine learning regression model and obtain the voltage mapping model is as follows: using the multidimensional color feature vector as the model input and the voltage of the single insulator corresponding to the effective emitting ROI as the model input. As labels, a machine learning regression model g(·) is trained. In this step, a random forest regression model is selected, with 100 trees and a fixed random seed to improve stability under small sample calibration conditions and reduce the risk of overfitting. The trained random forest regression model is saved as a voltage mapping model, with the expression being: = (fi), where, The voltage predicted by the model is the voltage prediction value of the i-th insulator output by the voltage mapping model.

[0051] In another embodiment, the machine learning regression model may also be support vector regression, K-nearest neighbor regression, gradient boosting tree, XGBoost (extreme gradient boosting tree), LightGBM (lightweight gradient boosting machine), CatBoost (categorical gradient boosting tree), extreme random tree, neural network, convolutional neural network, physically constrained neural network, Gaussian process regression, and piecewise calibration curve. The model output may be voltage, electric field strength, relative potential difference, and health score.

[0052] In another embodiment, the known equivalent electric field strength Ei can be used as a label, and a multidimensional color feature vector can be used as the model input. The machine learning regression model g(·) is trained through the above training process. This machine learning regression model is a random forest regression model, resulting in the electric field mapping model, i.e. = (fi), where, To predict the equivalent electric field for the model.

[0053] During the model training phase, cross-validation, leave-one-out validation, and group validation by camera model can also be performed to verify whether the trained model has sufficient stability under different shooting conditions. The model can simultaneously output mean absolute error, mean square error, and coefficient of determination. In addition, segmented error statistics for different voltage ranges are collected to provide a basis for setting accuracy thresholds during field deployment.

[0054] S5. The voltage prediction value of the insulator sheet is output through the voltage mapping model, and then the voltage prediction value of the insulator string is obtained. Based on the voltage prediction value of the insulator string, it is determined whether the effective emitting ROI is a valid image or an invalid image. If it is an invalid image, it is determined that the image in the bounding box corresponding to the effective emitting ROI is an invalid image or the voltage mapping model output is abnormal. If it is a valid image, proceed to step S6.

[0055] In this embodiment, a voltage mapping model is used to estimate the effective emitting region of interest (ROI) to obtain the voltage estimate of the insulator string. This process involves inputting the multidimensional color feature vector of the effective emitting ROI into the voltage mapping model to obtain the voltage estimate of the insulator sheet corresponding to the effective emitting ROI. Then, the estimated voltage values ​​of all insulator discs belonging to a single insulator string are summed to obtain the estimated voltage value of the insulator string. This step normalizes the estimated voltage value of the insulator string so that the sum of the normalized voltage values ​​of each insulator disc satisfies the total voltage constraint of the entire string.

[0056] In this embodiment, the specific process of determining whether the image in the bounding box corresponding to the effective ROI is a valid or invalid image based on the voltage prediction of the insulator string is as follows: the number of insulator pieces detected in the insulator string is n, and the preset anomaly threshold is... The expression for the voltage estimate of kV, n-piece insulators is: Where j is the summation index, ranging from 1 to n, here j=n, and n is the number of insulator discs in the insulator string. Let be the estimated voltage value for the j-th insulator.

[0057] like ≤ If the voltage is kV, then the effective ROI is determined to be an invalid image, and the image within the bounding box corresponding to the effective ROI is either an invalid image or the voltage mapping model output is abnormal; if > If the value is kV, the effective emitting ROI is determined to be a valid image, and step S6 is performed. This step couples the machine learning image estimation results with known physical quantities of the power system, which can offset the overall scale drift caused by changes in shooting conditions, coating batches, or camera parameters.

[0058] S6. Based on the voltage prediction of the insulator piece and the normal average voltage value of the insulator string, calculate the voltage ratio of the insulator piece corresponding to the effective emitting ROI, and obtain the grading result of each insulator piece in the image to be detected according to the voltage ratio.

[0059] In this embodiment, the process of calculating the voltage ratio of each insulator disc is as follows: S601. Calculate the normal equipotential voltage of the insulator string. The normal average voltage value of the insulator string is obtained by dividing the total voltage of the insulator string by the number of insulator discs in the insulator string. .

[0060] Normal equipotential voltage of insulator strings The calculation formula is: = / n, where, denoted as , where is the total voltage across the insulator string, and n is the number of insulator discs in the string.

[0061] S602. Calculate the normalized voltage of the insulator disc after the total voltage constraint of the insulator string. .

[0062] The calculation formula is: U i =U total ·û i / ( +ε), where û i Let û be the voltage prediction value of the i-th insulator output by the voltage mapping model. j Let be the voltage prediction value output by the voltage mapping model for the j-th insulator. Let j be the sum of the estimated voltage values ​​for n insulators, where j is the summation index, taking values ​​from 1 to n. = ε is a small constant to prevent the denominator from being zero, n is the number of insulator discs in the insulator string, and U total This represents the total voltage across the insulator string.

[0063] In another embodiment, at least one of the following can be introduced: insulator string end voltage distribution prior, fitting end enhancement coefficient, and adjacent plate smoothing coefficient, to normalize the voltage. Then, a second constraint update is performed. For example, in ultra-high voltage long-string insulators, the insulators near the high-voltage end usually bear higher voltages. The voltage prediction can be slightly corrected using a known prior distribution template to enhance the compatibility between lines of different voltage levels.

[0064] In another embodiment, the voltage U that depends only on the total voltage of the insulator string can be used. total The normalization method is replaced by the voltage ratio of adjacent insulator discs and the voltage ratio r of each disc. i The substitution quantities include the mean-variance standardized value, the historical baseline comparison value for the same tower and phase, the model prediction confidence interval discrimination value, the Bayesian anomaly detection value, the multi-frame voting results, and the time series trend analysis value. These substitution quantities are all applied to the normalized voltage sequence or state determination stage to improve robustness under different lines and shooting conditions.

[0065] For multiple frames of images of the same string of insulators, the normalized voltage U of the i-th insulator piece in each frame is... i Performing weighted fusion, the final fusion voltage of the i-th insulator can be written as: i =(Σ t q i (t) ·U i (t) ) / (Σ t q i (t) +ε), where, i Let q be the multi-frame fused voltage of the i-th insulator. i (t) U represents the weight of the fusion voltage of the i-th insulator in the t-th frame. i (t) Let be the voltage value of the i-th insulator in the t-th frame. This method can further suppress single-frame blurring, local occlusion, and instantaneous exposure changes in electroluminescent images, improving robustness in mobile inspection scenarios.

[0066] S603, based on the normalized voltage U i and normal average pressure value U avg The ratio is used to determine the voltage ratio r of the insulator disc. i .

[0067] In this embodiment of the application, based on the normal equalization voltage U of the insulator string avg and normalized voltage U i (If it is a single frame image, then it is U) i Multi-frame images are i Calculate the voltage ratio r of the insulator disc. i(Voltage ratio of the i-th insulator), and the specific calculation formula is: r i = U i / (U avg + ε).

[0068] S604. Obtain the grading results of each insulator sheet in the image to be detected according to the voltage ratio.

[0069] Grading each insulator sheet according to the voltage ratio r i includes comparing the voltage ratio r i with the first judgment threshold T1 and the second judgment threshold T2 to grade each insulator sheet.

[0070] If r i < T1, it is determined that the insulator sheet is a zero-value or severely deteriorated insulator; if T1 ≤ r i < T2, it is determined that the insulator sheet is a low-value deteriorated insulator; if r i ≥ T2, it is determined that the insulator sheet is a normal insulator; in this embodiment, T1 = 0.3 and T2 = 0.7. The thresholds T1 and T@ can be adjusted according to the insulator type, voltage level, string length, coating material, and operation and maintenance regulations.

[0071] Finally, based on the voltage ratio r i of the insulator sheet and the difference index d i between the insulator sheet and the adjacent insulator sheet, the adjacent sheet consistency verification of the grading result of the insulator sheet is performed to further judge the credibility of the grading result of the insulator sheet.

[0072] Specifically, the difference index d i reflects the local mutation degree of the insulator sheet relative to the adjacent insulator sheet.

[0073] d i = |U i (U i-1 + U i+1 ) / 2| / (U avg + ε), where U i is the normalized voltage of the determined insulator sheet, U i-1 is the normalized voltage of the insulator sheet adjacent to one side of the determined insulator sheet, and U i+1 is the normalized voltage of the insulator sheet adjacent to the other side of the determined insulator sheet.

[0074] In this embodiment, if r i < T2 and d i ≥ D H, it indicates that the voltage ratio of the i-th insulator disc is low and there is an obvious local mutation relative to the adjacent insulator discs. The grading result of this insulator disc has a high credibility. Subsequently, combined with the comprehensive credibility q of this insulator disc i and the credibility threshold q min to judge the grading result of this insulator disc.

[0075] If r i <T2 but d i <D L , it means that although the i-th insulator disc is judged as abnormal by r i , the difference from the adjacent insulator discs is not obvious. The grading result of this insulator disc can be marked as pending re-inspection or combined with the comprehensive credibility q i and judged again.

[0076] If D L ≤d i <D H , then maintain the grading result of this insulator disc obtained by r i , and mark its credibility as medium. Then, according to the comprehensive credibility q of this insulator disc i and the credibility threshold q min to further judge the grading result of this insulator disc.

[0077] If r i ≥T2 and d i <D L , then the grading result of this insulator disc in the normal state has a high credibility. Subsequently, combined with the comprehensive credibility q of this insulator disc i and the credibility threshold q min to judge the grading result of this insulator disc.

[0078] If r i ≥T2 but d i ≥D H , it means that there is a risk of local mutation in this insulator disc or the adjacent insulator discs, and a pending re-inspection mark should be output for the grading result of this insulator disc.

[0079] The above D L is preferably 0.15, and D H is preferably 0.30, and can be adjusted according to the insulator disc type, voltage level, string length, and historical inspection data.

[0080] The above process of judging the grading result of this insulator disc by combining the comprehensive credibility q of the insulator disc i and the credibility threshold q min is as follows: When q i ≥q minWhen q is reached, the grading result of the insulator sheet is taken as the final result and output directly; when q is reached... i min Instead of directly outputting the grading result of the insulator disc, the grading result of the insulator disc is marked as pending re-inspection or low confidence result.

[0081] q i =α·c det,i +β·ρ i +γ·( i / 255), where α is the detection confidence weight, β is the weight of the ratio of the effective luminous ROI area to the bounding box area, and γ is the weight of the effective luminous ROI saturation, preferably α=0.4, β=0.3, and γ=0.3; c det,i ρ represents the confidence level of the i-th bounding box. i The ratio of the effective ROI area to the bounding box area; i The mean saturation of the effective luminescent ROI; q min The confidence threshold is preferably set to 0.5, but can be adjusted based on the false positive and false negative rates on site.

[0082] In the insulator disc grading results, when two or more consecutive adjacent insulator discs on the same insulator string are judged as abnormal, and their adjacent disc difference indicators show obvious grouping characteristics, a high-level alarm can be triggered, indicating a local string segment abnormality or suggesting manual review of the overall coating condition of the string segment. Figure 3 As shown, it is designed to adapt to complex working conditions such as local contamination of insulator strings, local shading, or continuous deterioration of multiple insulator discs.

[0083] When outputting the graded status of each insulator disc, a table or distribution map is used for presentation. The table displays the location, voltage estimate, normalized voltage ratio, relative resistance, status, confidence level, and recommendations for each insulator disc, and generates a text diagnostic report. The distribution map is a voltage distribution bar chart with a normal voltage averaging reference line overlaid on it, facilitating maintenance personnel to quickly locate abnormal insulator discs. Furthermore, the diagnostic report can also output the relative deviation between the abnormal insulator disc and adjacent insulator discs, the voltage dispersion of the insulator disc string, a comparison of historical same-tower and same-phase inspection results, a list of suspected re-inspection insulator disc locations, and recommended maintenance priorities. This expands the output from simple algorithm results to an analytical tool for maintenance decision-making.

[0084] In addition, this application also provides an insulator grading system based on electroluminescent images. The system is implemented based on an insulator grading method using electroluminescent images, and includes: ​The image detection module is used to acquire the image to be detected. It detects the image to be detected through the insulator detection model to obtain the bounding box of each insulator piece in the insulator string. The order of the bounding boxes corresponds to the position order of the insulator pieces. The image to be detected is the electroluminescent image of the insulator string. The first judgment module is used to cluster the image pixels within the bounding box using a clustering algorithm to obtain a primary luminous image, process the primary luminous image to obtain candidate luminous areas, and select the largest connected component among the candidate luminous areas as the final luminous image. Based on the area ratio of the final luminous image within the corresponding bounding box, it is determined whether the effective luminous information of the final luminous image is sufficient; If insufficient, the final luminescent image is determined to have low confidence and is marked as requiring re-inspection; If sufficient, the final luminous image is determined to be an effective luminous ROI; a color space mapping transformation is performed on the effective luminous ROI to obtain a multidimensional color feature vector of the effective luminous ROI; the multidimensional color feature vector is used as the model input, and the voltage of the insulator sheet corresponding to the effective luminous ROI is used as the label to train a machine learning regression model to obtain a voltage mapping model. The output judgment module is used to output the voltage prediction value of the insulator piece corresponding to the effective emitting ROI through the voltage mapping model, thereby obtaining the voltage prediction value of the insulator string, and judging whether the effective emitting ROI is a valid image or an invalid image based on the voltage prediction value of the insulator string and a preset abnormal threshold. If the image is invalid, it is determined that the image in the bounding box corresponding to the effective ROI is invalid or the voltage mapping model output is abnormal. If it is a valid image, based on the voltage prediction of the insulator piece and the normal average voltage value of the insulator string, the voltage ratio of the insulator piece corresponding to the valid ROI is calculated, and the grading result of each insulator piece in the image to be detected is obtained according to the voltage ratio.

[0085] In addition, such as Figure 3 As shown, this embodiment also provides application examples of the above-mentioned zero-value insulator determination system based on electroluminescent images: Example: The object to be tested is an insulator string on a 110kV line, which has five insulator discs. First, the system generates five bounding boxes, which are then sorted vertically to obtain the bounding boxes for disc positions 1 to 5. Subsequently, Lab clustering and morphological cleansing are performed within each bounding box to obtain five effective luminous areas of interest (ROIs). After obtaining the voltage prediction using a voltage mapping model, U... total=110kV normalization is performed. If the voltage ratio r3 of the third insulator is significantly lower than 0.3, while the voltage ratios r2 and r4 of its adjacent insulators are close to 1, the system can classify the third insulator as a zero-value or severely degraded insulator and highlight it synchronously in the image and report.

[0086] It should be understood that the above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit them. Those skilled in the art can modify the technical solutions described in the above embodiments, or make equivalent substitutions for some of the technical features; and all such modifications and substitutions should fall within the protection scope of the appended claims of the present invention.

Claims

1. A method for classifying insulators based on electroluminescent images, characterized in that, include: The image to be detected is obtained, and the image is detected by the insulator detection model to obtain the bounding box of each insulator piece in the insulator string. The order of the bounding boxes corresponds to the position order of the insulator pieces. The image to be detected is the electroluminescent image of the insulator string. Clustering algorithms are used to cluster the image pixels within the bounding box to obtain a primary luminous image. The primary luminous image is then processed to obtain candidate luminous areas. The largest connected component among the candidate luminous areas is selected as the final luminous image. Based on the area ratio of the final luminous image within the corresponding bounding box, it is determined whether the effective luminous information of the final luminous image is sufficient; If insufficient, the final luminescent image is determined to have low confidence and is marked as requiring re-inspection; If sufficient, the final luminous image is determined to be an effective luminous ROI; a color space mapping transformation is performed on the effective luminous ROI to obtain a multidimensional color feature vector of the effective luminous ROI; the multidimensional color feature vector is used as the model input, and the voltage of the insulator sheet corresponding to the effective luminous ROI is used as the label to train a machine learning regression model to obtain a voltage mapping model. The voltage mapping model outputs the voltage estimate of the insulator piece corresponding to the effective ROI, thereby obtaining the voltage estimate of the insulator string. Based on the voltage estimate of the insulator string and a preset anomaly threshold, the effective ROI is determined to be a valid image or an invalid image. If the image is invalid, it is determined that the image in the bounding box corresponding to the effective ROI is invalid or the voltage mapping model output is abnormal. If it is a valid image, based on the voltage prediction of the insulator piece and the normal average voltage value of the insulator string, the voltage ratio of the insulator piece corresponding to the valid ROI is calculated, and the grading result of each insulator piece in the image to be detected is obtained according to the voltage ratio.

2. The insulator grading method based on electroluminescent images according to claim 1, characterized in that, The step of using a clustering algorithm to cluster the image pixels within the bounding box to obtain a primary luminous image, and processing the primary luminous image to obtain candidate luminous areas, includes: Clustering algorithms are used to cluster the image pixels within the bounding box to determine multiple cluster centers. The comprehensive chromaticity value of each cluster center is calculated, and the cluster center corresponding to the largest comprehensive chromaticity value is determined as the target cluster center. The primary luminous image is then determined based on the target cluster center. The primary luminescent image is filled with small holes and broken areas, and isolated noise is removed. The outer contour of the primary luminescent image is extracted, and the candidate luminescent area is determined based on the outer contour.

3. The insulator grading method based on electroluminescent images according to claim 1, characterized in that, The determination of whether the effective luminescence information of the final luminescence image is sufficient based on the area ratio of the final luminescence image within the corresponding bounding box includes: Compare the area ratio with the preset threshold; If the area ratio is less than a preset threshold, it is determined that the luminous information of the final luminous image is insufficient; If the area ratio is greater than or equal to a preset threshold, the final luminous image is determined to have sufficient luminous information.

4. The insulator grading method based on electroluminescent images according to claim 1, characterized in that, The effective luminous ROI is subjected to a color space mapping transformation to obtain a multidimensional color feature vector of the effective luminous ROI, including: The effective emitting ROI is subjected to a color space mapping transformation to map it to the HSV color space; The pixel distribution of the effective emitting ROI in the H channel, S channel and V channel of the HSV color space is statistically analyzed respectively; Based on the statistical pixel distribution results, a multidimensional color feature vector of the effective luminous ROI is established.

5. The insulator grading method based on electroluminescent images according to claim 1, characterized in that, The voltage mapping model outputs the voltage prediction value of the insulator sheet corresponding to the effective ROI, thereby obtaining the voltage prediction value of the insulator string, including: The multidimensional color feature vector of the effective emitting ROI is input into the voltage mapping model, and the voltage prediction value of the insulator corresponding to the effective emitting ROI is output. The estimated voltage values ​​of all the insulator discs on the insulator string are added together to obtain the estimated voltage value of the insulator string.

6. The insulator grading method based on electroluminescent images according to claim 1, characterized in that, Based on the voltage prediction of the insulator string and a preset anomaly threshold, the determination of whether the effective emitting ROI is a valid image or an invalid image includes: If the voltage estimate is less than or equal to a preset abnormal threshold, the effective ROI is determined to be an invalid image; If the voltage estimate is greater than the preset abnormal threshold, the effective ROI is determined to be a valid image.

7. The insulator grading method based on electroluminescent images according to claim 1, characterized in that, Based on the voltage prediction of the insulator piece and the normal voltage equalization value of the insulator string, the voltage ratio of the insulator piece corresponding to the effective ROI is calculated, including: The normal equal voltage value of the insulator string is determined by the ratio of the total voltage of the insulator string to the number of insulator discs in the insulator string. Calculate the normalized voltage U of the insulator disc after the total voltage constraint of the insulator string. i The calculation formula is: U i =U total ·û i / ( +ε); among them, U total The total voltage across the insulator string is û i Let û be the voltage prediction value of the i-th insulator piece output by the voltage mapping model. j Let be the voltage prediction value of the j-th insulator piece output by the voltage mapping model, n be the number of insulator pieces in the insulator string, ε be a small constant, and j be the summation index; The voltage ratio of the insulator disc is obtained by comparing the normalized voltage with the normal average voltage.

8. The insulator grading method based on electroluminescent images according to claim 1, characterized in that, The grading results for each insulator disc in the image to be detected are obtained based on the voltage ratio, including: The voltage ratio is compared with the first judgment threshold and the second judgment threshold; If the voltage ratio is less than the first judgment threshold, the insulator is determined to be zero or severely degraded. If the first judgment threshold is less than or equal to the voltage ratio and less than the second judgment threshold, the insulator is determined to be a low-value deteriorated insulator. If the voltage ratio is greater than or equal to the second judgment threshold, the insulator is determined to be a normal insulator.

9. An insulator grading system based on electroluminescent images, said system being implemented based on the method of any one of claims 1-8, characterized in that, The system includes: The image detection module is used to acquire the image to be detected. It detects the image to be detected through the insulator detection model to obtain the bounding box of each insulator piece in the insulator string. The order of the bounding boxes corresponds to the position order of the insulator pieces. The image to be detected is the electroluminescent image of the insulator string. The first judgment module is used to cluster the image pixels within the bounding box using a clustering algorithm to obtain a primary luminous image, process the primary luminous image to obtain candidate luminous areas, and select the largest connected component among the candidate luminous areas as the final luminous image. Based on the area ratio of the final luminous image within the corresponding bounding box, it is determined whether the effective luminous information of the final luminous image is sufficient; If insufficient, the final luminescent image is determined to have low confidence and is marked as requiring re-inspection; If sufficient, the final luminous image is determined to be an effective luminous ROI; a color space mapping transformation is performed on the effective luminous ROI to obtain a multidimensional color feature vector of the effective luminous ROI; the multidimensional color feature vector is used as the model input, and the voltage of the insulator sheet corresponding to the effective luminous ROI is used as the label to train a machine learning regression model to obtain a voltage mapping model. The output judgment module is used to output the voltage prediction value of the insulator piece corresponding to the effective emitting ROI through the voltage mapping model, thereby obtaining the voltage prediction value of the insulator string, and judging whether the effective emitting ROI is a valid image or an invalid image based on the voltage prediction value of the insulator string and a preset abnormal threshold. If the image is invalid, it is determined that the image in the bounding box corresponding to the effective ROI is invalid or the voltage mapping model output is abnormal. If it is a valid image, based on the voltage prediction of the insulator piece and the normal average voltage value of the insulator string, the voltage ratio of the insulator piece corresponding to the valid ROI is calculated, and the grading result of each insulator piece in the image to be detected is obtained according to the voltage ratio.

10. An electronic device, characterized in that, The device includes: At least one processor, and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor to enable the at least one processor to perform the method of any one of claims 1-8.

Images