Converter valve partial discharge ultraviolet image identification method based on SAM segmentation large model
By using a large model based on SAM segmentation, combined with small sample data expansion algorithms and quantization features, the automatic identification and precise location of partial discharge in converter valves were achieved. This solved the problems of manual reliance and insufficient model generalization in existing technologies, and improved the identification accuracy and equipment safety.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-27
- Publication Date
- 2026-03-24
AI Technical Summary
Existing technologies rely on human experience in the identification of partial discharge in converter valves, which is highly subjective, has a high false alarm rate, and traditional segmentation models have insufficient generalization ability, making it difficult to adapt to complex scenarios, resulting in inconsistent judgments and high maintenance costs.
A method based on the SAM segmentation model is adopted. By using a small sample data expansion algorithm and the SAM segmentation model, fine segmentation of the ultraviolet images of partial discharge is performed. Combined with quantization features, a partial discharge identification model is constructed to achieve automatic identification and accurate localization.
It improves the accuracy and consistency of partial discharge identification, reduces the manual workload, and ensures the safety and reliability of converter valve equipment.
Smart Images
Figure CN121725239A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of power equipment monitoring technology, and in particular to a method for recognizing partial discharge ultraviolet images of converter valves based on a large SAM segmentation model. Background Technology
[0002] As a core component of high-voltage direct current (HVDC) transmission systems, the insulation condition of converter valves directly affects the safe and stable operation of converter stations and the entire HVDC transmission project. During long-term operation, converter valves are susceptible to partial discharges on the valve tower surface and electrode edges due to factors such as electric field distortion, local defects, and contamination. If these partial discharges develop to a certain extent, they can cause insulation aging, localized ablation, and even converter valve shutdown. Currently, ultra-high sensitivity ultraviolet (UV) imaging devices are commonly used on-site to inspect the valve hall environment. By recording the UV photon radiation generated by partial discharges, the location, shape, and brightness of bright spots in the UV images are manually observed, and experience is combined to determine whether a discharge has occurred and its approximate location. However, this method has the following problems: 1. Relies on human experience and is highly subjective: Different maintenance personnel have different sensitivities and interpretation standards for ultraviolet bright spots. The same image may lead to different conclusions among different people, making it difficult to form a unified and objective basis for judgment.
[0003] 2. Ultraviolet images have complex backgrounds, resulting in a high false alarm rate: The valve hall is densely packed with equipment and contains interference sources such as reflective elements, stray light, and background noise. Traditional methods based on threshold segmentation or simple filtering are difficult to reliably distinguish between "real partial discharge bright spots" and "background noise bright spots", which can easily lead to false alarms or missed alarms.
[0004] 3. Traditional segmentation models lack generalization ability: Conventional dedicated object detection / segmentation networks require repeated labeling and training for specific scenarios. Once the layout of the equipment, the installation position of the ultraviolet camera, or the shooting distance changes, the algorithm performance is prone to significant degradation, resulting in high engineering maintenance costs.
[0005] With the development of large-scale visual foundational models, the Transformer-based segmentation model SAM has advantages such as strong versatility, insensitivity to target shape, and automatic generation of high-quality masks through prompts. It provides a new technical approach for fine segmentation in complex scenes. If the SAM segmentation model can be introduced into the UV image processing flow of partial discharge in converter valves, and combined with the prior knowledge of the converter valve structure, high-precision segmentation of photon distribution in UV images can be achieved, which is expected to realize automatic identification and accurate localization of partial discharge, reduce the manual burden and improve the consistency of judgment. Summary of the Invention
[0006] In view of this, the present invention proposes a method for partial discharge ultraviolet image recognition of converter valves based on a large visual segmentation model. By performing fine segmentation of the partial discharge ultraviolet image through a large visual segmentation model, the automatic detection of partial discharge is realized, providing a reliable image-based diagnostic means for defect monitoring of valve hall equipment.
[0007] The technical solution of this invention is implemented as follows: A method for recognizing partial discharge in converter valves using ultraviolet images based on a large SAM segmentation model includes the following steps: Step S1: Collect partial discharge ultraviolet image data including the converter valve, and construct a structured partial discharge ultraviolet image database; Step S2: Use a small sample data expansion algorithm to expand the images in the structured partial discharge ultraviolet image database; Step S3: Input the partial discharge ultraviolet image into the SAM segmentation model, and obtain the segmentation mask of the partial discharge region by the SAM segmentation model. Step S4: Extract the area, shape, and brightness of the partial discharge region as quantization features based on the segmentation mask, remove abnormal masks based on the quantization features, and obtain a set of quantization features; Step S5: Construct a partial discharge identification model based on quantization features, perform partial discharge ultraviolet image identification on the converter valve through the partial discharge identification model, and output whether there is partial discharge and the discharge level.
[0008] Preferably, step S1 includes the following steps: Step S1.1: Install an ultraviolet imaging device in the valve hall of the converter valve. Under different conditions such as different operating voltages, loads, and electromagnetic environments of the converter valve, collect ultraviolet images or video frames containing partial discharge of the converter valve equipment, and archive them by time, equipment number and operating condition information. Step S12: Divide the partial discharge ultraviolet images into normal, suspected discharge, and obvious discharge, select representative image samples, outline the bright spot region of partial discharge using the annotation tool, generate annotation masks, and store the partial discharge ultraviolet images, labels, and annotation masks together to form a structured partial discharge ultraviolet image database.
[0009] Preferably, step S2 includes the following specific steps: Step S2.1: Perform geometric transformations on the partial discharge ultraviolet image, including random rotation, scaling, translation, and mirror flipping; Step S2.2: Perform brightness perturbation on the partial discharge ultraviolet image, wherein the brightness perturbation includes adjusting the brightness, contrast and gamma value; Step S2.3: Inject noise into the partial discharge ultraviolet image, wherein the noise injection includes superimposed Gaussian noise and salt-and-pepper noise; Step S2.4: Perform pseudo-bright spot synthesis on the partial discharge ultraviolet image. The pseudo-bright spot synthesis generates pseudo-partial discharge bright spots based on a two-dimensional Gaussian function or an empirical bright spot template and superimposes them onto the partial discharge ultraviolet image. Step S2.5: Add the partial discharge ultraviolet image after geometric transformation, brightness perturbation, noise injection, and pseudo-bright spot synthesis to the structured partial discharge ultraviolet image database.
[0010] Preferably, step S3 includes the following specific steps: Step S3.1: If the partial discharge ultraviolet image is a single-channel grayscale image, convert it into a three-channel image by channel copying or pseudo-color mapping, while preserving the original spatial resolution or scaling it proportionally according to computing resources. Step S3.2: Combine the approximate location of partial discharge or the outline of the converter valve to construct the prompt information, and input the processed partial discharge ultraviolet image and prompt information into the SAM segmentation large model. The SAM segmentation large model will then output several candidate masks. Step S3.3: Sort the candidate masks according to the overlap between the candidate mask and the prompt area, average brightness and area, and select the candidate mask with the highest score as the segmentation mask for the partial discharge area.
[0011] Preferably, the specific steps of step S3.2 are as follows: Let the input partial discharge ultraviolet image be denoted as:
[0012] The prompt message is recorded as The SAM segmentation model first extracts features from the partial discharge ultraviolet image using an image encoder:
[0013] in, Represents an image coding network. These are the features of the downsampled image; The prompt information is embedded through a prompt encoder:
[0014] in, To prompt the encoder, To suggest embedded features; Mask decoder based on image features With hints and embedded features Interact with the mask to obtain the logits tensor:
[0015] in, Indicates mask decoder, This is the multi-channel mask logits tensor after upsampling to the original resolution. The number of candidate masks.
[0016] For the A candidate mask, which is in pixels The probability of a discharge region at a given location can be expressed as:
[0017] in, For logits, in Position corresponds to the The output value of the mask, It is the Sigmoid activation function. This indicates the predicted probability that the pixel belongs to a partial discharge region; For multi-scale masks at the same location, based on mask area, average brightness, or model confidence, from The optimal mask is selected as the segmentation result for that region.
[0018] Preferably, step S4 includes the following specific steps: Step S4.1: Extract area data, shape data, and brightness data of the partial discharge region as quantization features based on the segmentation mask; Step S4.2: Combine the pre-set area threshold, brightness threshold and position range to filter and remove abnormal masks, and combine the quantization features of the remaining segmentation masks into a quantization feature set.
[0019] Preferably, the area data includes area and equivalent radius, the shape data includes aspect ratio and roundness, and the brightness data includes average brightness and maximum brightness.
[0020] Preferably, step S5 includes the following specific steps: Step S5.1: Using quantified features as input and whether partial discharge occurs and the discharge level as output, a shallow neural network is used to construct a partial discharge identification model. Step S5.2: After acquiring the actual partial discharge ultraviolet image data, extract the quantitative features and input them into the partial discharge identification model. The partial discharge identification model will output whether there is a partial discharge and the discharge level.
[0021] Preferred options also include: Step S6: When the local discharge intensity or the cumulative number of occurrences exceeds the preset alarm threshold, an alarm message is generated and sent to the monitoring backend or the maintenance personnel terminal to indicate that the corresponding converter valve has a risk of partial discharge.
[0022] Preferably, the ultraviolet imaging device includes an ultrasensitive ultraviolet camera.
[0023] Compared with the prior art, the beneficial effects of the present invention are: The present invention provides a method for partial discharge ultraviolet image recognition of converter valves based on a large SAM segmentation model. After acquiring partial discharge ultraviolet images containing converter valve equipment, a small sample data expansion algorithm is introduced to generate diverse partial discharge ultraviolet images, solving the problem of limited discharge sample quantity. Then, a large SAM segmentation model is introduced to extract segmentation masks. Even in complex valve hall backgrounds, high-quality candidate bright spot masks can be generated, significantly improving the precision and robustness of partial discharge region segmentation, enhancing the accuracy and practical applicability of the partial discharge recognition model, and ensuring the safety of converter valve equipment. Quantization features can be extracted from the segmentation mask, and then the partial discharge identification model is trained based on the quantization features. After training, the partial discharge identification model can be put into use. The actual partial discharge ultraviolet images collected in the valve hall are processed by the SAM segmentation model to obtain the segmentation mask, and then quantization features are extracted. The quantization features are then used as the input of the partial discharge identification model. The partial discharge identification model can process and output whether there is partial discharge and the discharge level, thereby issuing alarm information to push to the monitoring backend or the terminal of the operation and maintenance personnel for timely processing. Attached Figure Description
[0024] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only preferred embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0025] Figure 1 The flowchart shows the method for recognizing partial discharge ultraviolet images of converter valves based on a large SAM segmentation model according to the present invention. Figure 2 This is a technical roadmap of the ultraviolet image recognition method for partial discharge of converter valves based on a large SAM segmentation model according to the present invention. Figure 3 This is a schematic diagram of the SAM segmentation result of a converter valve partial discharge ultraviolet image, representing an embodiment of the present invention's method for recognizing converter valve partial discharge based on a large SAM segmentation model. Detailed Implementation
[0026] To better understand the technical content of this invention, a specific embodiment is provided below, and the invention will be further described in conjunction with the accompanying drawings.
[0027] See Figures 1 to 3 The present invention provides a method for recognizing partial discharge ultraviolet images of converter valves based on a large SAM segmentation model, comprising the following steps: Step S1: Collect partial discharge ultraviolet image data including the converter valve, and construct a structured partial discharge ultraviolet image database; Step S2: Use a small sample data expansion algorithm to expand the images in the structured partial discharge ultraviolet image database; Step S3: Input the partial discharge ultraviolet image into the SAM segmentation model, and obtain the segmentation mask of the partial discharge region by the SAM segmentation model. Step S4: Extract the area, shape, and brightness of the partial discharge region as quantization features based on the segmentation mask, remove abnormal masks based on the quantization features, and obtain a set of quantization features; Step S5: Construct a partial discharge identification model based on quantization features, perform partial discharge ultraviolet image identification on the converter valve through the partial discharge identification model, and output whether there is partial discharge and the discharge level.
[0028] During operation, converter valves are susceptible to partial discharge due to external influences. When this partial discharge develops to a certain extent, it can cause aging of the valve tower insulation, localized ablation, or even valve shutdown. Therefore, accurate identification of partial discharge in converter valves is crucial. Common identification methods require acquiring UV images of partial discharges within the converter valve. However, due to the complexity of the scene and the difficulty of segmentation, the inability to perform fine segmentation can affect the identification of partial discharges. Therefore, this invention introduces a large-scale SAM segmentation model to achieve fine segmentation of UV images of partial discharges. After acquiring UV image data of partial discharges containing converter valves, a structured UV image database of partial discharges is constructed for backup. Since the number of partial discharge samples is limited, a small sample data expansion algorithm is used to expand the image samples in the structured UV image database of partial discharges. The expanded UV images of partial discharges can be input into the large-scale SAM segmentation model for candidate segmentation, obtaining images with fine contours and multi-scale features. The partial discharge region segmentation mask is obtained, and then quantization features are extracted from the segmentation mask. The quantization features can be used to identify abnormal masks, reducing the number of candidate masks, and can also be used as input for training the partial discharge identification model. The output of the partial discharge identification model is whether there is partial discharge and the discharge level. The trained partial discharge identification model can be used for partial discharge identification of converter valves. During the actual operation of the converter valve, actual partial discharge ultraviolet images can be collected, and after processing in steps S1-S4, the corresponding quantization features can be obtained. After the quantization features are input into the partial discharge identification model, the partial discharge identification model can output whether there is partial discharge and the discharge level. This invention no longer relies on a small sample dedicated network trained for a single scene. It can also generate high-quality candidate masks in complex valve hall backgrounds, significantly improving the precision and robustness of partial discharge region segmentation, thereby improving the accuracy and practical applicability of the partial discharge identification model and ensuring the safety of converter valve equipment.
[0029] Preferably, step S1 includes the following steps: Step S1.1: Install an ultraviolet imaging device in the converter valve hall. The ultraviolet imaging device includes an ultra-sensitive ultraviolet camera. Under different conditions such as different operating voltages, loads, and electromagnetic environments of the converter valve, collect ultraviolet images or video frames containing partial discharge of the converter valve equipment, and archive them by time, equipment number and operating condition information. Step S12: Divide the partial discharge ultraviolet images into normal, suspected discharge, and obvious discharge, select representative image samples, outline the bright spot region of partial discharge using the annotation tool, generate annotation masks, and store the partial discharge ultraviolet images, labels, and annotation masks together to form a structured partial discharge ultraviolet image database.
[0030] The preferred ultraviolet imaging device is an ultra-sensitive ultraviolet camera that achieves a sensitive response to ultraviolet radiation in the 240nm–280nm band through lens filtering and electron multiplication mechanisms. The field of view of the ultra-sensitive ultraviolet camera covers the entire converter valve tower or valve group area. The ultraviolet imaging device acquires partial discharge ultraviolet images of the converter valve under different operating conditions. The partial discharge ultraviolet images are tagged and managed according to equipment number, monitoring time, and operating condition information. Then, the partial discharge ultraviolet images are manually annotated. The images are classified into categories such as normal, suspected discharge, and obvious discharge through manual interpretation or experimental records. Representative image samples are selected to delineate the outline of the bright spot area of partial discharge and generate corresponding annotation masks. Finally, the partial discharge ultraviolet images, tags, and annotation masks are stored together to form a structured partial discharge ultraviolet image database, providing basic data support for subsequent data expansion and model recognition.
[0031] Preferably, step S2 includes the following specific steps: Step S2.1: Perform geometric transformations on the partial discharge ultraviolet image, including random rotation, scaling, translation, and mirror flipping; Step S2.2: Perform brightness perturbation on the partial discharge ultraviolet image, wherein the brightness perturbation includes adjusting the brightness, contrast and gamma value; Step S2.3: Inject noise into the partial discharge ultraviolet image, wherein the noise injection includes superimposed Gaussian noise and salt-and-pepper noise; Step S2.4: Perform pseudo-bright spot synthesis on the partial discharge ultraviolet image. The pseudo-bright spot synthesis generates pseudo-partial discharge bright spots based on a two-dimensional Gaussian function or an empirical bright spot template and superimposes them onto the partial discharge ultraviolet image. Step S2.5: Add the partial discharge ultraviolet image after geometric transformation, brightness perturbation, noise injection, and pseudo-bright spot synthesis to the structured partial discharge ultraviolet image database.
[0032] Based on the aforementioned partial discharge ultraviolet image database, and considering the limited number of typical discharge samples in actual engineering, a small-sample data expansion is performed on the labeled images. To address the limited number of labeled partial discharge samples, a small-sample data expansion algorithm is designed. On one hand, geometric enhancements such as random rotation, scaling, translation, and mirror flipping are applied to the partial discharge ultraviolet images to simulate different installation angles and viewing angle deviations. On the other hand, illumination enhancement is performed by adjusting brightness, contrast, and gamma values to simulate different exposure and background illumination conditions. Furthermore, Gaussian noise and salt-and-pepper noise are superimposed on the background area to simulate potential imaging noise in the field. Finally, pseudo-partial discharge bright spots can be generated at key locations on the equipment based on a two-dimensional Gaussian function or empirical bright spot templates and superimposed onto the partial discharge ultraviolet image, while simultaneously generating corresponding pseudo-label masks. This expands the partial discharge samples to different sizes, brightness levels, and morphologies, forming a diverse extended dataset that is then added to the structured partial discharge ultraviolet image database.
[0033] Preferably, step S3 includes the following specific steps: Step S3.1: If the partial discharge ultraviolet image is a single-channel grayscale image, convert it into a three-channel image by channel copying or pseudo-color mapping, while preserving the original spatial resolution or scaling it proportionally according to computing resources. Step S3.2: Combine the approximate location of partial discharge or the outline of the converter valve to construct the prompt information, and input the processed partial discharge ultraviolet image and prompt information into the SAM segmentation large model. The SAM segmentation large model will then output several candidate masks. Step S3.3: Sort the candidate masks according to the overlap between the candidate mask and the prompt area, average brightness and area, and select the candidate mask with the highest score as the segmentation mask for the partial discharge area.
[0034] Before being input into the SAM segmentation model, the partial discharge ultraviolet image needs to be adjusted according to the input requirements of the SAM segmentation model. When the partial discharge ultraviolet image is a single-channel grayscale image, it can be converted into a three-channel image through processing, and the resolution can be adjusted adaptively.
[0035] Then, combining the algorithm detection results, clue information can be constructed by combining the approximate location of partial discharge or the outline of the converter valve. The processed partial discharge ultraviolet image and clue information are then input into the SAM segmentation model. The SAM segmentation model outputs several candidate masks. In the specific implementation process, the candidate mask generation process of the SAM segmentation model is as follows: Let the input partial discharge ultraviolet image be denoted as:
[0036] The prompt information (including points, boxes, or prior masks) is denoted as The SAM segmentation model first extracts features from the partial discharge ultraviolet image using an image encoder:
[0037] in, Indicates an image encoder. These are the features of the downsampled image; The prompt information is embedded through a prompt encoder:
[0038] in, To prompt the encoder, To suggest embedded features; Mask decoder based on image features With hints and embedded features Interact with the mask to obtain the logits tensor:
[0039] in, Indicates mask decoder, This is the multi-channel mask logits tensor after upsampling to the original resolution. The number of candidate masks.
[0040] For the A candidate mask, which is in pixels The probability of a discharge region at a given location can be expressed as:
[0041] in, For logits, in Position corresponds to the The output value of the mask, It is the Sigmoid activation function. This indicates the predicted probability that the pixel belongs to a partial discharge region; For multi-scale masks at the same location, based on mask area, average brightness, or model confidence, from The optimal mask is selected as the segmentation result for that region.
[0042] For the large-scale SAM segmentation model, it mainly includes an image encoder, a cue encoder, and a mask decoder. The processed partial discharge ultraviolet image and cue information are input into the image encoder and cue encoder respectively for feature extraction. Then, the mask decoder interacts with the image features and cue embedding features to obtain the mask logits tensor and determine the candidate mask.
[0043] Furthermore, for any candidate mask, its position at the pixel level can be determined. The probability of the discharge region at a given location is calculated. Then, for multi-scale masks at the same location, the optimal mask can be selected based on the mask area, average brightness, or model confidence level to obtain the segmentation result for that region.
[0044] Finally, the candidate masks need to be sorted. Based on indicators such as the overlap between the candidate mask and the cue region, average brightness, and region area, the candidate mask with the highest score is used as the segmentation mask for the partial discharge region. Furthermore, morphological opening and closing operations can be combined to remove isolated noise points and smooth the segmentation boundary.
[0045] Preferably, step S4 includes the following specific steps: Step S4.1: Extract area data, shape data, and brightness data of the partial discharge region as quantitative features based on the segmentation mask. The area data includes area and equivalent radius, which are used to characterize the spatial scale of the discharge region. The shape data includes aspect ratio and roundness, which are used to distinguish different forms of discharge. The brightness data includes average brightness and maximum brightness, which are used to characterize the discharge intensity level. Step S4.2: Combine the pre-set area threshold, brightness threshold and position range to filter and remove abnormal masks, such as isolated noise with too small area and false alarm areas with too low brightness, so as to ensure that the final set of partial discharge ultraviolet quantization parameters has good stability and reliability. Combine the quantization features of the remaining segmented masks into a quantization feature set.
[0046] Preferably, step S5 includes the following specific steps: Step S5.1: Using quantified features as input and whether partial discharge occurs and the discharge level as output, a shallow neural network is used to construct a partial discharge identification model. Step S5.2: After acquiring the actual partial discharge ultraviolet image data, extract the quantitative features and input them into the partial discharge identification model. The partial discharge identification model will output whether there is a partial discharge and the discharge level.
[0047] The partial discharge identification model is trained using a shallow neural network. Its input is the quantified features obtained after processing the partial discharge ultraviolet image data through the SAM segmentation model and extracting features. The output is whether a discharge has occurred and the discharge level. The trained partial discharge identification model can be directly put into use. The actual partial discharge ultraviolet image data of the converter valve can be processed in steps S1-S4 to extract quantified features, which are then input into the partial discharge identification model. The partial discharge identification model outputs whether a partial discharge has occurred and the discharge level, realizing automatic identification of partial discharge based on the partial discharge ultraviolet image data. The discharge level includes severity levels such as slight discharge and obvious discharge.
[0048] Preferred options also include: Step S6: When the local discharge intensity or the cumulative number of occurrences exceeds the preset alarm threshold, an alarm message is generated and sent to the monitoring backend or the maintenance personnel terminal to indicate that the corresponding converter valve has a risk of partial discharge.
[0049] The number and intensity of partial discharges in the converter valve are statistically analyzed. When the intensity or cumulative occurrence of a partial discharge exceeds a preset alarm threshold, an alarm message is automatically generated and sent to the monitoring backend or maintenance personnel's terminal to provide a partial discharge risk warning. The alarm message includes a timestamp, device information, segmentation result image, and key feature parameters. Simultaneously, the corresponding original ultraviolet image, SAM segmentation mask, and judgment result are stored in the partial discharge event database for subsequent trend analysis, defect tracing, and algorithm optimization.
[0050] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A method for partial discharge ultraviolet image recognition of converter valves based on a large SAM segmentation model, characterized in that, Includes the following steps: Step S1: Collect partial discharge ultraviolet image data including the converter valve, and construct a structured partial discharge ultraviolet image database; Step S2: Use a small sample data expansion algorithm to expand the images in the structured partial discharge ultraviolet image database; Step S3: Input the partial discharge ultraviolet image into the SAM segmentation model, and obtain the segmentation mask of the partial discharge region by the SAM segmentation model. Step S4: Extract the area, shape, and brightness of the partial discharge region as quantization features based on the segmentation mask, remove abnormal masks based on the quantization features, and obtain a set of quantization features; Step S5: Construct a partial discharge identification model based on quantization features, perform partial discharge ultraviolet image identification on the converter valve through the partial discharge identification model, and output whether there is partial discharge and the discharge level.
2. The method for partial discharge ultraviolet image recognition of converter valves based on a large SAM segmentation model according to claim 1, characterized in that, The specific steps of step S1 include: Step S1.1: Install an ultraviolet imaging device in the valve hall of the converter valve. Under different conditions such as different operating voltages, loads, and electromagnetic environments of the converter valve, collect ultraviolet images or video frames containing partial discharge of the converter valve equipment, and archive them by time, equipment number and operating condition information. Step S12: Divide the partial discharge ultraviolet images into normal, suspected discharge, and obvious discharge, select representative image samples, outline the bright spot region of partial discharge using the annotation tool, generate annotation masks, and store the partial discharge ultraviolet images, labels, and annotation masks together to form a structured partial discharge ultraviolet image database.
3. The method for partial discharge ultraviolet image recognition of converter valves based on a large SAM segmentation model according to claim 1, characterized in that, The specific steps of step S2 include: Step S2.1: Perform geometric transformations on the partial discharge ultraviolet image, including random rotation, scaling, translation, and mirror flipping; Step S2.2: Perform brightness perturbation on the partial discharge ultraviolet image, wherein the brightness perturbation includes adjusting the brightness, contrast and gamma value; Step S2.3: Inject noise into the partial discharge ultraviolet image, wherein the noise injection includes superimposed Gaussian noise and salt-and-pepper noise; Step S2.4: Perform pseudo-bright spot synthesis on the partial discharge ultraviolet image. The pseudo-bright spot synthesis generates pseudo-partial discharge bright spots based on a two-dimensional Gaussian function or an empirical bright spot template and superimposes them onto the partial discharge ultraviolet image. Step S2.5: Add the partial discharge ultraviolet image after geometric transformation, brightness perturbation, noise injection, and pseudo-bright spot synthesis to the structured partial discharge ultraviolet image database.
4. The method for partial discharge ultraviolet image recognition of converter valves based on a large SAM segmentation model according to claim 1, characterized in that, The specific steps of step S3 include: Step S3.1: If the partial discharge ultraviolet image is a single-channel grayscale image, convert it into a three-channel image by channel copying or pseudo-color mapping, while preserving the original spatial resolution or scaling it proportionally according to computing resources. Step S3.2: Combine the approximate location of partial discharge or the outline of the converter valve to construct the prompt information, and input the processed partial discharge ultraviolet image and prompt information into the SAM segmentation large model. The SAM segmentation large model will then output several candidate masks. Step S3.3: Sort the candidate masks according to the overlap between the candidate mask and the prompt area, average brightness and area, and select the candidate mask with the highest score as the segmentation mask for the partial discharge area.
5. The method for partial discharge ultraviolet image recognition of converter valves based on a large SAM segmentation model according to claim 4, characterized in that, The specific steps of step S3.2 are as follows: Let the input partial discharge ultraviolet image be denoted as: The prompt message is recorded as The SAM segmentation model first extracts features from the partial discharge ultraviolet image using an image encoder: in, Represents an image coding network. These are the features of the downsampled image; The prompt information is embedded through a prompt encoder: in, To prompt the encoder, To suggest embedded features; Mask decoder based on image features With hints and embedded features Interact with the mask to obtain the logits tensor: in, Indicates mask decoder, This is the multi-channel mask logits tensor after upsampling to the original resolution. The number of candidate masks. For the A candidate mask, which is in pixels The probability of a discharge region at a given location can be expressed as: in, For logits, in Position corresponds to the The output value of the mask, It is the Sigmoid activation function. This indicates the predicted probability that the pixel belongs to a partial discharge region; For multi-scale masks at the same location, based on mask area, average brightness, or model confidence, from The optimal mask is selected as the segmentation result for that region.
6. The method for partial discharge ultraviolet image recognition of converter valves based on a large SAM segmentation model according to claim 1, characterized in that, The specific steps of step S4 include: Step S4.1: Extract area data, shape data, and brightness data of the partial discharge region as quantization features based on the segmentation mask; Step S4.2: Combine the pre-set area threshold, brightness threshold and position range to filter and remove abnormal masks, and combine the quantization features of the remaining segmentation masks into a quantization feature set.
7. The method for partial discharge ultraviolet image recognition of converter valves based on a large SAM segmentation model according to claim 6, characterized in that, The area data includes area and equivalent radius, the shape data includes aspect ratio and roundness, and the brightness data includes average brightness and maximum brightness.
8. The method for partial discharge ultraviolet image recognition of converter valves based on a large SAM segmentation model according to claim 1, characterized in that, The specific steps of step S5 include: Step S5.1: Using quantified features as input and whether partial discharge occurs and the discharge level as output, a shallow neural network is used to construct a partial discharge identification model. Step S5.2: After acquiring the actual partial discharge ultraviolet image data, extract the quantitative features and input them into the partial discharge identification model. The partial discharge identification model will output whether there is a partial discharge and the discharge level.
9. The method for partial discharge ultraviolet image recognition of converter valves based on a large SAM segmentation model according to claim 1, characterized in that, Also includes: Step S6: When the local discharge intensity or the cumulative number of occurrences exceeds the preset alarm threshold, an alarm message is generated and sent to the monitoring backend or the maintenance personnel terminal to indicate that the corresponding converter valve has a risk of partial discharge.
10. The method for recognizing partial discharge ultraviolet images of converter valves based on a large SAM segmentation model according to claim 2, characterized in that, The ultraviolet imaging device includes an ultrasensitive ultraviolet camera.