Insulator temperature detection method and system

Abstract

The application discloses a kind of insulator temperature detection method and system, wherein method includes obtaining the visible light image of target insulator, infrared light image and multi-source environmental parameter;The visible light image and infrared light image are registered and handled, and double light registration image is obtained;Double light registration image is pixel-level segmentation, and insulator effective area image is generated according to segmentation result;Insulator effective area image is extracted, and visible light modal feature and infrared light modal feature are obtained, and multi-source environmental parameter is based on visible light modal feature and infrared light modal feature Feature fusion is carried out, and cross-modal fusion feature is obtained;Temperature prediction is carried out based on cross-modal fusion feature, and first temperature prediction result is obtained;First temperature prediction result is corrected based on multi-source environmental parameter, and the temperature detection result of target insulator is obtained.The method of the application improves the efficiency of insulator temperature detection.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to an insulator temperature detection method and system. Background Technology

[0002] Insulators are core components in high-voltage / ultra-high-voltage transmission lines, ensuring insulation performance and operational safety. Their condition directly affects the stable operation of the power system. Zero-value insulators, due to insulation failure, can cause abnormal local temperature increases, potentially leading to serious accidents such as flashover and tripping. Therefore, accurate temperature monitoring of zero-value insulators is a critical task in power operation and maintenance.

[0003] Traditional methods of insulator temperature detection mainly rely on manual observation from towers or inspection with telescopes, which are not only inefficient and labor-intensive, but also pose safety hazards under high-voltage environments.

[0004] Existing intelligent detection technologies can detect insulators through image recognition, such as locating and detecting the temperature of target insulators using single visible or infrared light. However, single visible light detection can only complete the initial location of the insulator and cannot identify zero-value insulators without obvious defects based on appearance features. While single infrared light detection can identify zero values ​​based on temperature anomaly signals, infrared images generally suffer from low resolution and blurred edges. In complex backgrounds such as conductors, metal supports, and vegetation, the positioning accuracy is poor, and it is easily affected by environmental factors such as sunlight radiation and high-temperature reflection from metal components, resulting in large temperature detection errors. Consequently, the accuracy and anti-interference capabilities of the detection results are insufficient to meet the actual needs of automated operation and maintenance of high-voltage transmission lines. Summary of the Invention

[0005] This invention provides an insulator temperature detection method and system to solve the technical problem of how to improve existing insulator temperature detection methods and achieve the effect of improving the efficiency of insulator temperature detection.

[0006] To address the aforementioned technical problems, the present invention provides an insulator temperature detection method, comprising: Acquire visible light images, infrared images, and multi-source environmental parameters of the target insulator; The visible light image and the infrared light image are registered to obtain a dual-light registered image; The dual-light registration image is segmented at the pixel level, and an effective insulator region image is generated based on the segmentation results; wherein, the pixel-level segmentation is designed to separate the insulator from the background based on the insulator texture features in the visible light image and the insulator temperature range features in the infrared light image. Feature extraction is performed on the image of the effective area of ​​the insulator to obtain visible light mode features and infrared light mode features. Based on the multi-source environmental parameters, the visible light mode features and the infrared light mode features are fused to obtain cross-modal fusion features. Temperature prediction is performed based on the cross-modal fusion features to obtain a first temperature prediction result; The first temperature prediction result is corrected based on multi-source environmental parameters to obtain the temperature detection result of the target insulator.

[0007] As one preferred embodiment, the registration processing of the visible light image and the infrared light image to obtain a dual-light registered image includes: Based on the typical color and geometric features of insulators, the region of interest of the insulator in the visible light image is located. Based on the region of interest of the insulator, feature points are extracted from the visible light image and the infrared light image respectively to obtain visible light feature points and infrared light feature points; The visible light feature points and the infrared light feature points are matched using a feature matching algorithm to generate matching feature point pairs. The pixel-level mapping relationship between the visible light image and the infrared light image is determined based on the matching feature point pairs, and the image transformation is performed on the visible light image or the infrared light image based on the pixel-level mapping relationship to generate a dual-light registration image.

[0008] As one preferred embodiment, the step of performing pixel-level segmentation on the dual-light registration image and generating an effective insulator region image based on the segmentation results includes: Histogram equalization is performed on the first infrared image in the dual-light registration image, and the insulator temperature range feature is extracted from the obtained second infrared image to obtain the insulator temperature range feature. Insulator texture features are extracted from the first visible light image in the dual-light registration image to obtain insulator texture features; The dual-light registration image is classified at the pixel level based on the insulator temperature range characteristics and the insulator texture characteristics. The pixel region corresponding to the target insulator is determined according to the pixel-level classification results to generate an effective insulator region image.

[0009] As one preferred embodiment, the feature fusion of the visible light modal features and the infrared light modal features based on the multi-source environmental parameters to obtain cross-modal fused features includes: The information contribution of the visible light modal features and the infrared light modal features is evaluated based on the multi-source environmental parameters, and the fusion weights of the visible light modal features and the infrared light modal features are generated according to the information contribution. The visible light modal features and the infrared light modal features are fused based on the fusion weights to obtain cross-modal fused features.

[0010] As one preferred embodiment, the step of correcting the first temperature prediction result based on multi-source environmental parameters to obtain the temperature detection result of the target insulator includes: Acquire historical multi-source environmental parameters and historical temperature data of the target insulator; A correlation analysis was performed on the historical multi-source environmental parameters and the historical temperature data to obtain the interference patterns of different environmental factors on temperature detection. Based on the aforementioned interference pattern, a temperature correction coefficient corresponding to the multi-source environmental parameters is generated; The first temperature prediction result is corrected according to the temperature correction coefficient to obtain the temperature detection result of the target insulator.

[0011] Another aspect of the present invention provides an insulator temperature detection system, comprising: The acquisition module is used to acquire visible light images, infrared images, and multi-source environmental parameters of the target insulator; The registration module is used to register the visible light image and the infrared light image to obtain a dual-light registered image; The segmentation module is used to perform pixel-level segmentation on the dual-light registration image and generate an effective insulator region image based on the segmentation results; wherein, the pixel-level segmentation is designed to separate the insulator from the background based on the insulator texture features in the visible light image and the insulator temperature range features in the infrared light image. The fusion module is used to extract features from the effective area image of the insulator to obtain visible light mode features and infrared light mode features, and to fuse the visible light mode features and infrared light mode features based on the multi-source environmental parameters to obtain cross-modal fusion features; The prediction module is used to predict temperature based on the cross-modal fusion features to obtain a first temperature prediction result; The generation module is used to correct the first temperature prediction result based on multi-source environmental parameters to obtain the temperature detection result of the target insulator.

[0012] As one preferred embodiment, the registration module is specifically used for: Based on the typical color and geometric features of insulators, the region of interest of the insulator in the visible light image is located. Based on the region of interest of the insulator, feature points are extracted from the visible light image and the infrared light image respectively to obtain visible light feature points and infrared light feature points; The visible light feature points and the infrared light feature points are matched using a feature matching algorithm to generate matching feature point pairs. The pixel-level mapping relationship between the visible light image and the infrared light image is determined based on the matching feature point pairs, and the image transformation is performed on the visible light image or the infrared light image based on the pixel-level mapping relationship to generate a dual-light registration image.

[0013] As one preferred embodiment, the segmentation module is specifically used for: Histogram equalization is performed on the first infrared image in the dual-light registration image, and the insulator temperature range feature is extracted from the obtained second infrared image to obtain the insulator temperature range feature. Insulator texture features are extracted from the first visible light image in the dual-light registration image to obtain insulator texture features; The dual-light registration image is classified at the pixel level based on the insulator temperature range characteristics and the insulator texture characteristics. The pixel region corresponding to the target insulator is determined according to the pixel-level classification results to generate an effective insulator region image.

[0014] As one preferred embodiment, the fusion module is specifically used for: The information contribution of the visible light modal features and the infrared light modal features is evaluated based on the multi-source environmental parameters, and the fusion weights of the visible light modal features and the infrared light modal features are generated according to the information contribution. The visible light modal features and the infrared light modal features are fused based on the fusion weights to obtain cross-modal fused features.

[0015] As one preferred embodiment, the correction module is specifically used for: Acquire historical multi-source environmental parameters and historical temperature data of the target insulator; A correlation analysis was performed on the historical multi-source environmental parameters and the historical temperature data to obtain the interference patterns of different environmental factors on temperature detection. Based on the aforementioned interference pattern, a temperature correction coefficient corresponding to the multi-source environmental parameters is generated; The first temperature prediction result is corrected according to the temperature correction coefficient to obtain the temperature detection result of the target insulator.

[0016] Compared with the prior art, the beneficial effects of the present invention are at least one of the following: 1) This invention effectively overcomes the limitations of single-mode detection by simultaneously acquiring visible light images, infrared images, and multi-source environmental parameters, combined with dual-light registration and pixel-level segmentation techniques. Pixel-level segmentation achieves precise separation of insulators from the background based on visible light texture features and infrared temperature range features, significantly reducing interference from complex backgrounds such as conductors, metal supports, and vegetation. Simultaneously, dual-light registration ensures spatial consistency of the two modal information, laying a high-precision foundation for subsequent feature fusion and temperature detection. Furthermore, dynamically adjusting the fusion weights of the dual-light modal features based on multi-source environmental parameters adaptively matches the information contribution of the two modes according to different environmental scenarios, fully leveraging the complementary advantages of accurate visible light positioning and sensitive infrared temperature measurement, and significantly improving the reliability of insulator positioning and feature extraction in complex environments.

[0017] 2) This invention predicts temperature by fusing features across modalities and then corrects the prediction results using multi-source environmental parameters, constructing a dual-precision temperature measurement mechanism of "feature fusion prediction + environmental parameter correction". This design effectively counteracts the interference of environmental factors such as solar radiation, metal reflection, and changes in air temperature and humidity on temperature measurement, significantly reducing temperature detection errors and ensuring the accuracy of the detection results. Simultaneously, the entire process automates image acquisition, preprocessing, feature extraction, and temperature output without manual intervention. This avoids the safety hazards of manual inspections and significantly improves detection efficiency. It is adaptable to the complex operation and maintenance scenarios of high-voltage / ultra-high-voltage transmission lines, providing reliable technical support for the rapid identification of zero-value insulators and the safe and stable operation of power systems. Attached Figure Description

[0018] Figure 1 This is a schematic flowchart of an insulator temperature detection method in one embodiment of the present invention; Figure 2 This is a structural block diagram of an insulator temperature detection system according to one embodiment of the present invention; Figure label: The module consists of: 11. Acquisition module; 12. Registration module; 13. Segmentation module; 14. Fusion module; 15. Prediction module; and 16. Generation module. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The purpose of providing these embodiments is to make the disclosure of the present invention more thorough and comprehensive. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0020] In the description of this invention, the terms "first," "second," "third," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined with "first," "second," "third," etc., may explicitly or implicitly include one or more of that feature. In the description of this invention, unless otherwise stated, "a plurality of" means two or more.

[0021] In the description of this invention, it should be noted that, unless otherwise defined, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art. The terminology used in this specification is for the purpose of describing specific embodiments only and is not intended to limit the invention. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0022] One embodiment of the present invention provides a method for detecting the temperature of an insulator. For details, please refer to [link to relevant documentation]. Figure 1 , Figure 1 The diagram shown is a flowchart of an insulator temperature detection method according to one embodiment of the present invention, which includes steps S1-S6: S1: Acquire visible light images, infrared images, and multi-source environmental parameters of the target insulator; Insulators, as core insulating components of high-voltage / ultra-high-voltage transmission lines, directly determine the operational safety of these lines due to their insulation performance. Localized temperature anomalies caused by insulation failure in zero-value insulators are a key contributing factor to line flashovers, tripping, and other accidents. Therefore, accurate detection of zero-value insulators is a core task of power operation and maintenance. Existing technologies for acquiring dual-light images and environmental parameters generally suffer from problems such as spatiotemporal asynchrony, limited parameter types, and insufficient equipment accuracy. This makes subsequent registration and fusion, as well as temperature detection, susceptible to environmental interference such as sunlight radiation and metal reflection, and the detection accuracy is insufficient to meet the requirements of automated operation and maintenance. Therefore, this step S1, through standardized acquisition equipment selection, strict acquisition condition limitations, and the design of a high-precision synchronization mechanism, acquires high-quality, spatiotemporally consistent visible light images, infrared images, and multi-source environmental parameters, laying a reliable data foundation for the accurate implementation of subsequent stages.

[0023] In this embodiment, a coaxial dual-light integrated camera is used as the core image acquisition device. This coaxial design ensures that the field of view for visible light and infrared light acquisition is completely consistent, reducing the difficulty and error of subsequent image registration from a hardware perspective. Preferably, the visible light resolution of the dual-light integrated camera is not less than 1920×1080 pixels, which can clearly capture the texture details of the insulator surface and provide sufficient visual information for insulator positioning and segmentation; the infrared light resolution is not less than 640×512 pixels, which can accurately present the temperature distribution of the insulator and meet the accuracy requirements for temperature anomaly identification; its temperature measurement range is set to -20℃~150℃, comprehensively covering the temperature change range of transmission line insulators under different climatic conditions such as severe cold and extreme heat, and the temperature measurement accuracy is controlled within ±0.5℃, providing a core guarantee for the reliability of the initial temperature data. It should be noted that this invention does not limit the specific brand and model of the dual-light integrated camera. Any commercially available integrated camera or customized camera that meets the above parameter requirements is applicable, such as an integrated device equipped with a high-definition CMOS visible light module and an uncooled infrared focal plane array infrared module.

[0024] To comprehensively quantify and counteract the interference of environmental factors on temperature detection, this embodiment integrates a multi-source environmental sensor with a dual-light integrated camera to achieve synchronous acquisition of environmental parameters. Specifically, the multi-source environmental sensor integrates at least three types of core sensors: 1) a solar radiation intensity sensor, with a measurement range of 0~1000W / m² and an accuracy of ±5W / m², used to collect the solar radiation intensity at the detection site in real time, providing data support for subsequent judgment of visible light mode reliability and temperature correction; 2) an air temperature and humidity sensor, with a temperature measurement accuracy of ±0.3℃ and a humidity measurement accuracy of ±3%RH, which corrects the influence of environmental temperature and humidity on infrared thermometry by capturing changes in air temperature and humidity; 3) a metal component temperature sensor, with a measurement range of -20℃~200℃ and an accuracy of ±0.4℃, specifically for collecting the surface temperature of surrounding metal components such as conductors and metal supports, used to distinguish between the temperature of the insulator itself and false temperature signals caused by high-temperature reflection from metal. Preferably, the multi-source environmental sensor adopts a modular design, and can flexibly add or remove expansion modules such as wind speed sensors and precipitation sensors according to the environmental complexity of the actual application scenario. This invention does not further limit this.

[0025] To ensure the validity and consistency of the collected data, this embodiment strictly regulates the key conditions during the acquisition process. Regarding the acquisition distance, the distance between the dual-light integrated camera and the target insulator is controlled within the range of 5-20m. This range ensures the clarity of the insulator image, avoiding feature blurring due to excessive distance, while also preventing mutual occlusion of insulator discs or incomplete field of view caused by excessively close proximity. Regarding the shooting angle, the angle between the camera's shooting direction and the plane of the insulator string should not exceed 30° to minimize occlusion between insulator discs and ensure complete imaging of the effective area of ​​each insulator disc. It should be noted that during the acquisition process, scenarios with no severe insulator occlusion should be prioritized. If occlusion cannot be avoided due to line layout limitations, it should be mitigated by adjusting the shooting position and angle or changing the acquisition point to reduce the difficulty of subsequent data preprocessing. If occlusion is truly unavoidable, the occlusion situation should be noted in the data labels to provide a reference for subsequent model inference.

[0026] Since the time difference between environmental parameters and dual-light image acquisition directly affects the accuracy of environmental correction, this embodiment uses a hardware synchronization module to achieve spatiotemporal consistency control of data acquisition. Specifically, the acquisition time difference between the visible light image and the infrared image of the dual-light integrated camera is ≤10ms, ensuring a high degree of matching between the spatial position and temperature state of the insulator corresponding to the two modes of images, avoiding registration deviations caused by acquisition time differences; the acquisition time difference between the parameters of the multi-source environmental sensor and the dual-light image acquisition is ≤5ms, ensuring that the environmental parameters can truly reflect the on-site environmental state at the moment of image acquisition, providing accurate environmental references for subsequent adaptive fusion and temperature correction. In this embodiment, the synchronization module adopts a clock signal hard triggering method, controlling the camera shutter and sensor sampling through a unified synchronization pulse, which has higher synchronization accuracy and stability compared to software synchronization methods; of course, under the premise of meeting the above time difference requirements, a hardware and software combined synchronization scheme can also be used, and this invention does not limit this.

[0027] To adapt to the complex operation and maintenance scenarios of high-voltage / ultra-high-voltage transmission lines, the dual-light integrated camera and multi-source environmental sensor in this embodiment can be mounted on different types of mobile carriers. Preferably, for large-scale, long-distance transmission line inspections, the data acquisition equipment can be mounted on a drone to achieve long-distance, contactless automated data acquisition, significantly reducing the labor intensity of maintenance personnel and the safety risks under high-voltage environments. For detailed inspections of local lines and areas difficult for drones to reach, such as mountainous regions, the data acquisition equipment can be integrated into a handheld terminal to improve the flexibility of data acquisition operations. It should be noted that the choice of carrier must ensure the stability of the data acquisition equipment to avoid image blurring or sensor data fluctuations caused by carrier shaking. For example, drones need to have a hovering accuracy of ≤0.5m, and handheld terminals need to be equipped with image stabilization modules and convenient mounting brackets to ensure the quality of the acquired data.

[0028] S2: Perform registration processing on the visible light image and the infrared light image to obtain a dual-light registered image; In the dual-light fusion detection of zero-value insulators, the registration accuracy of the dual-light images directly determines the reliability of subsequent feature fusion and temperature detection. Existing technologies often employ global feature point matching for dual-light registration, failing to address the insulator region specifically. This results in an excessive number of invalid feature points in complex backgrounds such as conductors, vegetation, and metal supports, easily leading to mismatches and making registration errors difficult to control. Consequently, this causes problems such as feature distortion and positioning deviations after fusion. Therefore, this step optimizes the entire process from "insulator ROI pre-positioning - ROI feature point extraction - precise feature matching - pixel-level mapping transformation" to achieve high-precision registration of visible light and infrared images, providing spatially aligned dual-light image data for subsequent pixel-level segmentation and cross-modal fusion.

[0029] In this embodiment, the Region of Interest (ROI) of the insulator in the visible light image is first located based on the typical color and geometric features of the insulator, thereby narrowing the feature extraction range and reducing background interference. Preferably, the HSV color space thresholding method is used to achieve ROI pre-location, where the Hue value ranges from 0 to 30°, the Saturation value ranges from 0 to 100, and the Value value ranges from 100 to 255. This threshold range is based on statistics from a large number of insulator samples and can accurately filter out common gray-white, light yellow, and other color regions of insulators. At the same time, combined with the geometric features of the insulator (such as long strips or string distribution), isolated background noise areas are removed through morphological dilation, erosion operations, and contour detection, and finally, the complete insulator string ROI is output. It should be noted that the specific algorithm for ROI location is not limited in this invention. If the insulator surface is dirty or stained, the RGB color space thresholding method or deep learning semantic segmentation method can also be used to achieve ROI location, as long as the insulator can be effectively separated from the background. Among them, "Region of Interest (ROI)" refers to a specific area in the image that has research value, specifically the area where the insulator is located. Its core function is to focus on key detection objects and reduce the computational complexity and interference of subsequent processing.

[0030] After the insulator ROI is located, feature point extraction is performed only within the ROI for both visible and infrared images to avoid background feature interference caused by global extraction. Preferably, an improved SIFT (Scale Invariant Feature Transform) algorithm is used for feature point extraction. This algorithm possesses scale invariance and rotation invariance, effectively addressing issues such as scale differences and shooting angle changes that may occur during dual-light image acquisition. Specifically, a feature point extraction threshold of 0.03 is set within the ROI, retaining only high-contrast feature points with response values ​​higher than this threshold to ensure good discriminative power of the extracted feature points. Simultaneously, orientation assignment and descriptor generation are performed on the extracted feature points, with each feature point corresponding to a 128-dimensional SIFT descriptor for subsequent feature matching. It should be noted that other scale-invariant feature extraction algorithms such as SURF and ORB can also be used in this invention, as long as they meet the registration accuracy requirements, they are all within the scope of protection of this invention. The SIFT algorithm is an image local feature extraction algorithm. It obtains stable feature points of an image through five steps: constructing a scale space, detecting extreme points, refining extreme points, assigning directions, and generating descriptors. It is suitable for image matching under different scales, rotations, and lighting conditions.

[0031] After feature point extraction, the visible light feature points and infrared light feature points are matched using the FLANN (Fast Nearest Neighbor) algorithm to generate initial matching feature point pairs. Preferably, the FLANN matching adopts the K-Nearest Neighbor matching strategy, setting K to 2. That is, for each visible light feature point, the two nearest feature points in the infrared light feature point set are searched as candidate matching points. Then, the "nearest neighbor distance ratio" criterion is used to screen valid matching points. If the distance ratio between the first and second nearest neighbors in the candidate matching points is less than 0.8, the matching point pair is retained; otherwise, it is judged as a mismatch and discarded. To further improve the matching accuracy, this embodiment introduces the RANSAC (Random Sample Consensus) algorithm to perform a secondary screening of the initial matching point pairs. The number of iterations is set to 1500 and the confidence threshold is 0.95. By randomly sampling a small number of matching point pairs, the initial mapping model is solved, and then based on this model, interior points that conform to the model (valid matching point pairs) are screened, while outliers (mismatched point pairs) are discarded. It should be noted that this invention does not limit the specific combination of feature matching algorithms and screening criteria. For example, a brute-force matching algorithm combined with RANSAC screening can also be used, as long as accurate screening of effective matching point pairs can be achieved. Among them, the "FLANN algorithm" is an efficient nearest neighbor search algorithm. Compared with traditional brute-force matching, it has a faster matching speed and is suitable for large-scale feature point matching scenarios. The "RANSAC algorithm" is a robust parameter estimation algorithm that can effectively remove outliers (mismatched points) in the data and improve the accuracy of model solution.

[0032] After obtaining valid matching feature point pairs, the Homography matrix is ​​solved based on these matching point pairs to determine the pixel-level mapping relationship between the visible light image and the infrared light image. The Homography matrix is ​​a 3×3 projection transformation matrix that describes the projection mapping relationship between two planar images. It can be solved using at least four pairs of non-collinear valid matching point pairs. In this embodiment, the least squares method is used to fit the valid matching point pairs to obtain the optimal Homography matrix. Subsequently, a perspective transformation is performed on the infrared light image based on this matrix to achieve pixel-level alignment between the insulator regions in the infrared light image and the insulator regions in the visible light image, ultimately generating a two-light registered image. It should be noted that a perspective transformation can also be performed on the visible light image. This invention does not limit the transformation object, as long as it can achieve spatial alignment of the insulator regions in the two-light image. At the same time, the registered two-light image must satisfy the positional deviation of corresponding pixels of the insulator sheet ≤ 1 pixel to ensure the accuracy of subsequent pixel-level segmentation and feature fusion. Among them, the "Homography matrix" is the core matrix in computer vision that describes the projection relationship between two planes. It can realize image transformations such as rotation, translation, scaling, and perspective, and is the key core of two-light image registration.

[0033] This embodiment narrows the feature extraction range by pre-locating the insulator ROI, reducing background interference from the source; it extracts stable feature points using an improved SIFT algorithm, and ensures matching accuracy by combining FLANN and RANSAC dual screening; it achieves pixel-level mapping transformation through the Homography matrix, ultimately controlling the registration error to within 1 pixel, which is significantly better than traditional registration methods. It should be noted that this invention does not absolutely limit the specific parameters in each step (such as the HSV threshold range, feature point extraction threshold, RANSAC iteration count, etc.). Those skilled in the art can make adaptive adjustments according to the actual application scenario (such as insulator type, shooting environment, image resolution, etc.). As long as high-precision registration of dual-light images can be achieved, it falls within the protection scope of this invention.

[0034] S3: Perform pixel-level segmentation on the dual-light registration image and generate an effective insulator region image based on the segmentation results; wherein, the pixel-level segmentation is designed to separate the insulator from the background based on the insulator texture features in the visible light image and the insulator temperature range features in the infrared light image. In the dual-light fusion detection process for zero-value insulators, accurate separation of the insulator from the background is a core prerequisite for subsequent feature extraction and temperature detection. Existing technologies often rely on single-mode features (such as visible light color or infrared temperature alone), which struggles to handle complex background interference. For example, high-temperature reflections from conductors and metal supports are easily misinterpreted as insulator temperature features, and the textures of vegetation and the sky are similar to those of the insulator, leading to segmentation confusion. This ultimately results in incomplete extraction of the effective insulator region and residual background noise, severely impacting the accuracy of subsequent temperature detection. Therefore, this step employs a complete workflow design of "infrared image enhancement - dual-mode feature collaborative extraction - pixel-level accurate classification." Leveraging the complementary advantages of visible light texture features and infrared temperature range features, it achieves pixel-level separation of the insulator from the background, generating a high-purity image of the effective insulator region, laying a solid foundation for subsequent cross-modal fusion and temperature detection.

[0035] In this embodiment, the first infrared image in the dual-light registration image is first subjected to histogram equalization to enhance the discriminability of the temperature range features of the insulator. Infrared images, limited by their imaging principle, generally suffer from low contrast and indistinct temperature gradients, especially when the temperature difference between the insulator and the background is small, making direct extraction of temperature features prone to missed detections. Preferably, an adaptive histogram equalization (CLAHE) algorithm is used. This algorithm divides the image into multiple sub-blocks (tileGridSize set to 8×8), performs histogram equalization on each sub-block individually, and sets a contrast limit parameter clipLimit=2.0 to avoid temperature feature distortion caused by local overexposure. If the temperature difference in the infrared image is extremely large (e.g., the temperature difference between the metal support and the insulator exceeds 30°C), a global histogram equalization (equalizeHist) algorithm can also be used to simplify the calculation process. It should be noted that this invention does not limit the specific histogram equalization algorithm; any algorithm that can enhance the temperature gradient of the infrared image and improve the discriminability of temperature features falls within the scope of protection of this invention. Among them, "histogram equalization" is an image enhancement technique that expands the dynamic range of an image by adjusting the distribution of gray values, making dark details clearer and bright details richer. Its core function here is to enhance the temperature difference characteristics between the insulator and the background.

[0036] After histogram equalization, a second infrared image is obtained, and the temperature range feature of the insulator is extracted based on this image. In this embodiment, the extraction of the insulator temperature range feature needs to be dynamically adjusted in conjunction with environmental parameters, rather than using a fixed threshold. First, the reasonable temperature range of the insulator is determined by using air temperature, solar radiation intensity, and metal component temperature obtained from multi-source environmental sensors: for example, when the air temperature is 25℃ and the solar radiation intensity is ≤300W / m², the normal temperature range of the insulator is set to 23℃~40℃; when the solar radiation intensity is ≥500W / m², considering the temperature rise caused by direct sunlight, the normal temperature range is adjusted upward to 28℃~48℃; when the temperature difference between the metal component and the air temperature is ≥15℃, the high temperature range corresponding to metal reflection (metal temperature ±3℃) is excluded to avoid misjudging the metal reflection temperature as the insulator temperature feature. Subsequently, each pixel of the second infrared image is traversed, the pixel temperature value is extracted, and it is determined whether it is within the dynamically adjusted insulator temperature range. Pixels within this range are marked as "temperature candidate pixels," and a temperature candidate pixel mask is generated. It should be noted that the reference values ​​for the above temperature range can be adaptively adjusted according to the type of insulator (such as porcelain, glass, or composite insulators) and voltage level. This invention does not impose an absolute limitation on this. The core is to correct the temperature threshold through environmental parameters to ensure the accuracy of the temperature range characteristics.

[0037] Simultaneously with the infrared temperature range feature extraction, the insulator texture features are extracted from the first visible light image in the dual-light registration image to capture the structural features of the insulator surface. The insulator surface typically has regular textures (such as stripes in porcelain insulators and wrinkles in composite insulators), while the background (conductors, metal supports, vegetation) often has irregular or no obvious texture. This difference is key to distinguishing the insulator from the background. Preferably, the gray-level co-occurrence matrix (GLCM) is used to extract texture features, and the specific steps are as follows: First, the first visible light image is converted into a grayscale image. Using a 5×5 sliding window and a 1-pixel step size, the grayscale co-occurrence matrix is ​​calculated in four directions: 0°, 45°, 90°, and 135°. From the matrix in each direction, four core features—contrast, correlation, energy, and entropy—are extracted to form a 16-dimensional texture feature vector (4 directions × 4 feature dimensions). Contrast reflects the clarity of the texture, correlation reflects the consistency of the texture, energy characterizes the uniformity of the texture, and entropy reflects the complexity of the texture. These four features work together to comprehensively characterize the texture properties of the insulator. Of course, this invention can also use other texture feature extraction algorithms such as Gabor filters and LBP (Local Binary Pattern), as long as they can effectively distinguish the texture differences between the insulator and the background, they are all within the scope of protection of this invention.

[0038] After acquiring the temperature range features (temperature candidate pixel mask) and texture features (16-dimensional feature vector) of the insulator, pixel-level classification of the dual-light registration image is achieved based on dual-feature fusion. Preferably, the MaskR-CNN algorithm is used as the core algorithm for pixel-level segmentation. This algorithm can simultaneously achieve object detection and pixel-level segmentation, and output accurate binary masks. In this embodiment, the input of the MaskR-CNN algorithm is the dual-light registration image, and the extracted texture feature vector and temperature candidate pixel mask are used as auxiliary inputs and integrated into the feature fusion layer of the algorithm. The training process of the algorithm is based on the pre-trained model on the COCO dataset, and fine-tuned on the insulator-specific dataset. The training objective is to optimize the pixel classification loss (cross-entropy loss) so that the model can accurately determine whether each pixel belongs to an insulator. The specific pixel-level classification process is as follows: For each pixel in the dual-light registration image, the model combines its corresponding texture feature vector and temperature value (whether it is within the temperature candidate range), and outputs a classification result of "insulator" or "background" through a fully connected layer, ultimately generating a binary mask. In the mask, regions with a pixel value of 1 correspond to insulators, and regions with a pixel value of 0 correspond to the background (wires, metal supports, vegetation, sky, etc.). It should be noted that this invention does not limit the pixel-level classification algorithm; deep learning algorithms with pixel-level segmentation capabilities, such as U-Net, SegNet, and DeepLab, are all applicable. The core is to achieve collaborative classification of dual-modal features. "Pixel-level segmentation" refers to assigning each pixel in the image to a specific category (such as insulator or background), achieving fine-grained separation of the target and background, with accuracy far exceeding that of traditional region-level segmentation.

[0039] To further improve the quality of the insulator effective area image, the generated binary mask undergoes post-processing optimization. First, a morphological erosion operation (using a 3×3 rectangular structuring element, iterating once) is applied to remove isolated noise points smaller than 50 pixels in the mask (such as tiny bright areas caused by direct sunlight or image sensor noise). Then, a morphological dilation operation (also using a 3×3 rectangular structuring element, iterating once) is performed to repair tiny gaps at the edges of the insulator area, ensuring the integrity of the effective insulator area. Finally, the optimized binary mask is multiplied pixel-by-pixel with the dual-light registration image, setting the pixels in the background area (mask value 0) to invalid values ​​(e.g., setting RGB images to black, infrared images to invalid temperature measurement values), retaining only the pixel information of the insulator area (mask value 1), thus generating the insulator effective area image. It should be noted that the shape (e.g., circle, rectangle), size, and number of iterations of the morphological operation structuring element can be adjusted according to the image noise level; this invention does not limit these adjustments, as long as noise removal and area repair are achieved.

[0040] This embodiment enhances infrared temperature features through histogram equalization, dynamically determines temperature range thresholds based on environmental parameters, captures insulator structural characteristics based on multi-dimensional texture features, and achieves accurate separation of the insulator from the background through a pixel-level segmentation algorithm with dual-feature fusion. The resulting effective insulator region image effectively eliminates background interference such as metal reflection, direct sunlight, and vegetation obstruction, achieving a segmentation accuracy of over 98% for the insulator region. This design not only overcomes the limitations of single-feature segmentation but also provides high-purity effective region data for subsequent cross-modal feature fusion and temperature detection, significantly improving the anti-interference capability and detection accuracy of the entire detection system.

[0041] S4: Extract features from the effective area image of the insulator to obtain visible light mode features and infrared light mode features, and fuse the visible light mode features and infrared light mode features based on the multi-source environmental parameters to obtain cross-modal fusion features; In the dual-light fusion detection of zero-value insulators, the completeness of feature extraction and the targeted nature of cross-modal fusion directly determine the accuracy of subsequent temperature prediction and zero-value identification. Existing technologies often employ a single network structure for dual-light modal feature extraction, failing to adequately adapt to the differences in characteristics between the two modes. Furthermore, feature fusion is often a simple superposition or fixed-weight summation, neglecting the dynamic influence of environmental parameters such as solar radiation and metal reflection on the contribution of the two modal information. For instance, in strong light environments, the visible light mode is susceptible to reflective interference; using equal weights for fusion in such cases leads to distorted fused features. Similarly, in high-temperature scenarios involving metal components, the infrared light mode is easily affected by background interference; fixed weights cannot enhance effective features, ultimately impacting detection accuracy. Therefore, this step employs a design of "dual-branch targeted feature extraction - environmental parameter-driven weight evaluation - adaptive weighted fusion" to achieve accurate fusion of visible and infrared light modal features, generating cross-modal fusion features with both positioning reliability and temperature sensitivity, providing high-quality feature support for subsequent temperature prediction.

[0042] In this embodiment, a two-branch feature extraction is first performed on the insulator effective area image to obtain visible light mode features and infrared light mode features respectively. Preferably, a YOLOv8-based two-branch feature extraction framework is used. The two input branches of this framework receive the segmented visible light effective area image and infrared light effective area image respectively. Each branch follows the original feature extraction structure of YOLOv8, including a C2f module and a SPPF (Spatial Pyramid Pooling Fusion) module to ensure the efficiency and multi-scale adaptability of feature extraction. For the visible light effective area image branch, the C2f module fully captures the texture details of the insulator surface (such as the striped texture of porcelain insulators and the surface wrinkles of composite insulators) through residual connections and multi-scale feature fusion. These texture features are the core basis for insulator localization and background differentiation. The SPPF module aggregates multi-scale texture features through pooling operations at different scales to improve the global consistency of features. For the infrared effective region image branch, the temperature distribution features of the insulators (such as local high-temperature regions and temperature gradient changes in zero-value insulators) are extracted using the C2f module. These features are then combined with the SPPF module to aggregate multi-scale temperature features, enhancing the feature representation of temperature anomaly regions. Finally, both branches output visible light mode feature maps. (256 channels) and infrared light mode feature map (The number of channels is 256). It should be noted that this invention does not limit the specific structure of the feature extraction network. Networks with strong feature expression capabilities, such as ResNet and EfficientNet, can also be used to replace the feature extraction module of YOLOv8, as long as they can adapt to the feature extraction requirements of the two modes respectively. Among them, "visible light mode features" refers to the texture, shape and other features extracted from visible light images for insulator localization and contour recognition, and "infrared light mode features" refers to the temperature distribution, gradient changes and other features extracted from infrared light images for temperature anomaly recognition.

[0043] After feature extraction, a modal confidence evaluation unit is constructed based on multi-source environmental parameters to dynamically evaluate the information contribution of the two modal features and generate corresponding fusion weights. The multi-source environmental parameters include solar radiation intensity, air humidity, and metal component temperature, which together determine the information reliability of the two modalities: solar radiation intensity directly affects the clarity of texture features in the visible light modality; air humidity affects the temperature detection stability of the infrared light modality; and metal component temperature relates to whether the infrared light modality is interfered with by high background temperatures. In this embodiment, the modal confidence evaluation unit consists of two fully connected layers, with an input dimension of 3 (corresponding to the three types of environmental parameters) and an output dimension of 2 (corresponding to the fusion weights of the two modalities). The specific evaluation logic is as follows: 1) When the solar radiation intensity is ≥500W / m², the visible light mode is more susceptible to reflected light interference, resulting in a decrease in information contribution. The evaluation unit outputs the visible light mode weights. ( The value is 0.3~0.5, and the infrared mode weight is... ( The value is 0.5~0.7; 2) When the solar radiation intensity is <500W / m², the visible light modal texture features are clear, the information contribution is improved, and the weight is adjusted to... ( =0.5~0.7 ( =0.3~0.5; 3) When the temperature difference between the metal component and the initial temperature measurement of the insulator is ≥10℃, the infrared light mode is easily affected by metal reflection. While outputting the weight, the evaluation unit triggers the background suppression factor of the infrared light feature (value 0.8~0.9) to further enhance the temperature feature of the insulator itself. 4) When the air humidity is ≥80%RH, the error in infrared mode temperature detection increases, so the weight of visible mode should be appropriately increased by 0.1~0.2. It should be noted that the above weight reference range is based on statistical analysis of a large amount of experimental data. Those skilled in the art can fine-tune it according to the insulator type and voltage level, and the network structure of the evaluation unit can also be replaced with CNN, Transformer, etc., as long as the mapping relationship between environmental parameters and fusion weights can be learned, all of which fall within the protection scope of this invention.

[0044] To ensure the effectiveness of feature fusion, the number of channels in the two modal feature maps needs to be unified first. Since there may be differences in the number of channels during the feature extraction process for different modalities (e.g., after replacing the feature extraction network), this embodiment uses a 1×1 convolution on the visible light modal feature map. infrared light mode feature map Channel number calibration is performed to unify the channel number of both to 256 dimensions, ensuring dimensionality consistency during weighting. The core function of 1×1 convolution is to achieve linear transformation of channel dimensions and feature fusion without changing the feature map space size, avoiding fusion distortion caused by channel number mismatch. It should be noted that if the number of channels in the output feature maps after dual-branch feature extraction is already consistent, this step can be omitted; this invention does not impose such a limitation.

[0045] After weight generation and channel unification are completed, the feature maps of the two modalities are adaptively weighted and fused based on the fusion weights to obtain cross-modal fused features. Preferably, the feature fusion is achieved using an adaptive weighted summation formula, as follows: in, For visible light mode fusion weights, For infrared mode fusion weights, It consists of a set of multi-source environmental parameters (solar radiation intensity, air humidity, and metal component temperature) and satisfies the constraints. This ensures energy conservation during the fusion process. The core advantage of this formula lies in its dynamic adjustment of the contribution ratios of the two modes based on environmental parameters. When a mode is less affected by environmental disturbances, it is given a higher weight, allowing its core features to dominate the fusion result; conversely, when a mode is more affected by disturbances, its weight is reduced to minimize the impact of noise on the fused features. For example, in a strong outdoor light environment (solar radiation intensity 650 W / m²), this can be achieved. =0.4, =0.6, the fusion feature focuses more on infrared light temperature characteristics to avoid visible light reflection interference; under cloudy conditions (solar radiation intensity 200W / m²). =0.6 =0.4, the fusion feature focuses more on visible light texture features, improving positioning accuracy.

[0046] To further optimize the expressive power of the fused features, this embodiment performs nonlinear activation and feature enhancement processing on the fused feature map. Preferably, a ReLU activation function is introduced after weighted summation to enhance the nonlinear expressive power of the fused features through nonlinear transformation and suppress invalid feature responses. Subsequently, a 3×3 convolution kernel is used for feature smoothing to reduce feature abrupt changes caused by weight switching and improve the spatial consistency of the fused features. It should be noted that the specific activation function and feature enhancement methods can be flexibly adjusted, such as using the GELU activation function, BatchNorm normalization, etc. As long as the optimization of the fused features can be achieved, it falls within the protection scope of this invention. Among them, "cross-modal fusion features" refers to the comprehensive features obtained by dynamically weighting and fusing visible light and infrared light modal features. It combines the positioning accuracy of visible light modes with the temperature sensitivity of infrared light modes and is the core basis for subsequent temperature prediction and zero-value insulator identification.

[0047] This embodiment fully leverages the core information of both modalities through dual-branch feature extraction and dynamically adjusts the fusion weights based on multi-source environmental parameters, achieving "environmentally adaptive" cross-modal fusion and effectively overcoming the limitations of traditional fixed-weight fusion. The generated cross-modal fusion features can flexibly adapt to the advantages of both modalities according to different environmental scenarios, maintaining high-quality feature expression even in complex environments. This lays a solid foundation for subsequent temperature prediction and zero-value identification based on fusion features, significantly improving the anti-interference capability and detection accuracy of the entire detection system.

[0048] S5: Based on the cross-modal fusion features, perform temperature prediction to obtain a first temperature prediction result; In the zero-value insulator temperature detection process, temperature prediction based on cross-modal fusion features is a core step connecting feature extraction and final temperature correction. Existing technologies often employ single regression layers or simple fully connected networks for temperature prediction, failing to fully explore the deep correlation between texture and temperature information within the cross-modal fusion features. Furthermore, they lack model optimization for accurate prediction of the insulator's center point temperature. This results in predictions that are easily affected by insufficient feature representation and inadequate model generalization ability, making it difficult to accurately reflect the insulator's true temperature state and consequently impacting the effectiveness of subsequent environmental correction. Therefore, this step optimizes the prediction network structure, designs targeted regression branches and loss functions, and fully leverages the complementary advantages of cross-modal fusion features to achieve accurate preliminary prediction of the insulator's center point temperature, providing a high-quality initial temperature prediction result for subsequent temperature correction.

[0049] In this embodiment, the temperature prediction network is integrated based on the existing YOLOv8 adaptive fusion framework to ensure the coherence of feature transfer and prediction efficiency. The cross-modal fusion feature map serves as the input to the prediction network. This feature map combines the localization accuracy of the visible light mode with the temperature sensitivity of the infrared light mode, with a unified 256-dimensional channel count and spatial dimensions corresponding to the input image (after feature extraction and fusion, it becomes a 640×640 pixel multi-scale feature map of 1 / 8 to 1 / 32 scale). It should be noted that this invention does not limit the specific scale of the fused feature map, as long as it covers the multi-scale features of the effective insulator area and ensures the spatial correlation of temperature prediction. The "cross-modal fusion feature" refers to the comprehensive feature map obtained through an adaptive weighted fusion algorithm, which integrates the core information of visible light texture features and infrared temperature features, serving as the core data foundation for temperature prediction.

[0050] To achieve accurate temperature regression at the center point, a dedicated temperature regression branch is added to the YOLOv8 detector. This branch is designed in parallel with the classification branch and the rotated box regression branch, sharing the multi-scale features of the cross-modal fusion feature map.

[0051] Preferably, the temperature regression branch adopts a structure of "multi-scale feature fusion + fully connected layer regression": First, the multi-scale features (such as 80×80, 40×40, 20×20) of the cross-modal fusion feature map are upsampled and concatenated. The number of channels is compressed to 128 dimensions through 1×1 convolution, reducing computational complexity while retaining key features. Subsequently, three fully connected layers are connected. The first fully connected layer has an input dimension of 128×H×W (H and W are the height and width of the concatenated feature map) and an output dimension of 512 dimensions. The second fully connected layer has an output dimension of 128 dimensions, and the third fully connected layer has an output dimension of 1 dimension, directly corresponding to the temperature prediction value of the insulator center point. BatchNorm normalization is added after each fully connected layer to accelerate model convergence, improve generalization ability, and avoid overfitting. It should be noted that the number of fully connected layers and the output dimension can be adjusted according to the size of the dataset and the required prediction accuracy. For example, when the dataset has a large sample size, it can be increased to 4 fully connected layers, and the output dimension can be adjusted to 256 dimensions and then compressed to 1 dimension. This invention does not impose an absolute limitation on this.

[0052] To adapt to the numerical characteristics of temperature prediction, the temperature regression branch uses the ReLU activation function for nonlinear transformation. The expression of the ReLU activation function is f(x) = max(0,x). Its core advantage lies in effectively alleviating the gradient vanishing problem while retaining the positive and effective information in cross-modal fusion features. Since the insulator temperature is a non-negative value, ReLU activation avoids interference from negative feature responses on temperature prediction, ensuring the reasonableness of the prediction results. In this embodiment, the ReLU activation function is applied after the first two fully connected layers. The third fully connected layer does not add an activation function, directly outputting a continuous numerical temperature prediction result, avoiding nonlinear distortion of the temperature value by the activation function. Of course, this invention can also use better activation functions such as GELU and Swish, as long as they meet the nonlinear expression requirements of temperature prediction and do not introduce invalid interference, they are all within the scope of protection of this invention. The "ReLU activation function" is a commonly used deep learning activation function that improves the nonlinear expression capability and training stability of the model by retaining positive inputs and suppressing negative inputs.

[0053] To ensure the training effectiveness of the temperature regression branch, the MSE (mean squared error) loss function is used as the loss calculation criterion for temperature prediction. This loss function effectively measures the squared difference between the predicted and true temperatures, is sensitive to temperature errors, and can help the model quickly learn an accurate temperature mapping relationship. The formula for calculating the MSE loss function is as follows: in, The number of training samples. For the first The actual center point temperature of each insulator disc (obtained by a high-precision contact thermometer with an accuracy of ±0.1℃). For the first The predicted temperature of each insulator piece is calculated. During model training, the temperature regression loss, the rotation box localization loss (CIoU loss), and the insulator classification loss (cross-entropy loss) constitute a multi-task loss function, with a loss weight ratio of 2.0:1.0:0.5. This prioritizes temperature prediction during training, ensuring that temperature prediction accuracy is optimized in conjunction with localization and classification accuracy. It should be noted that if imbalanced temperature anomaly samples exist in practical applications, a weighted MSE loss function can be used to assign higher weights to temperature anomaly samples such as zero-value insulators, improving the accuracy of anomaly temperature prediction. This invention does not limit this approach.

[0054] During the model inference phase, the temperature prediction process is performed simultaneously with the localization and classification inference to ensure detection efficiency. Specifically, after the cross-modal fusion feature map to be tested is input into the temperature regression branch, it undergoes multi-scale feature fusion, fully connected layer mapping, and ReLU activation processing. The third fully connected layer directly outputs the continuous numerical temperature of the center point of each insulator sheet, which is the first temperature prediction result. To ensure the reliability of the prediction results, reasonable numerical constraints are set on the predicted temperature during the inference process: based on the actual operating temperature range of the insulator (-20℃~150℃), predicted temperatures exceeding this range are set as invalid values ​​(removed or re-predicted during subsequent correction) to avoid interference from extreme outliers in subsequent processing. In this embodiment, the model inference time is ≤40ms / frame, and the error of the first temperature prediction result is controlled within ±0.8℃, providing accurate and reliable basic data for subsequent temperature correction based on multi-source environmental parameters.

[0055] This embodiment, through an integrated temperature regression branch design, multi-scale feature fusion, and targeted activation and loss function selection, fully mines the core temperature-related information in cross-modal fusion features, achieving efficient and accurate preliminary prediction of the temperature at the insulator's center point. This design ensures the synergy between temperature prediction, localization, and classification tasks, and improves prediction accuracy through a specialized network structure and training strategy. It effectively solves the problems of weak generalization ability and large errors in traditional temperature prediction models, laying a crucial foundation for the accuracy of the final temperature detection results.

[0056] S6: Correct the first temperature prediction result based on multi-source environmental parameters to obtain the temperature detection result of the target insulator.

[0057] In the zero-value insulator temperature detection process, although the initial temperature prediction result achieves preliminary accuracy based on cross-modal fusion features, it is still subject to residual interference from environmental factors such as solar radiation, air temperature and humidity, and high-temperature reflection from metal components. For example, direct sunlight can cause a false increase in the insulator surface temperature, high-temperature reflection from metal components can be misinterpreted as the insulator's own temperature, and changes in air humidity can affect the energy transfer efficiency of infrared thermometry. These factors can all lead to deviations between the initial temperature prediction result and the actual temperature, making it unsuitable as the final detection result. Existing technologies lack dynamic temperature correction mechanisms for multi-source environmental parameters, often employing fixed correction values ​​or simple linear corrections, which are difficult to adapt to complex and variable outdoor environments, resulting in insufficient temperature measurement accuracy. Therefore, this step uses historical data correlation analysis to uncover environmental interference patterns, generates dynamic correction coefficients, and accurately corrects the initial temperature prediction result, ensuring the accuracy and reliability of the final temperature detection result.

[0058] In this embodiment, historical multi-source environmental parameters and historical temperature data of the target insulator are first acquired to construct a historical dataset for analyzing environmental interference patterns. The historical multi-source environmental parameters are consistent with the real-time environmental parameters collected in S1, including solar radiation intensity (0~1000W / m²), air temperature (-20℃~50℃), air humidity (10%RH~95%RH), and metal component temperature (-20℃~200℃). All parameters are acquired using multi-source environmental sensors of the same specifications as those acquired in real-time to ensure data consistency. The historical temperature data is the actual temperature of the insulator's center point measured by a high-precision contact thermometer (accuracy ±0.1℃), serving as the benchmark for temperature error analysis. To ensure the comprehensiveness and generalization of interference patterns, the historical dataset needs to cover different application scenarios, including different voltage levels (110kV, 220kV, 500kV, 1000kV), different insulator types (porcelain, glass, composite insulators), different weather conditions (sunny, cloudy, partly cloudy, light rain), and different time periods (morning, noon, evening). At least 10,000 sets of valid historical data samples should be collected, with each sample containing "historical multi-source environmental parameters + corresponding actual insulator temperature + predicted first temperature of the same period (simulated generation)," constructing a three-dimensional correlated dataset. It should be noted that this invention does not impose an absolute limit on the specific sample size of the historical data, as long as it covers the main application scenarios and supports reliable mining of interference patterns. The sample size can be dynamically expanded according to actual application needs.

[0059] Multi-dimensional correlation analysis was performed on the constructed historical dataset to uncover the quantitative relationship between different environmental factors and temperature prediction errors, i.e., the environmental interference patterns. Preferably, a combined analysis method of "multiple linear regression + gradient boosting tree (XGBoost)" was adopted, where multiple linear regression was used to initially screen environmental factors that significantly affect temperature errors, and gradient boosting tree was used to construct a nonlinear interference model. The specific analysis process is as follows: 1) Define temperature error ΔT = first temperature prediction result - actual temperature as the dependent variable for correlation analysis; 2) Using the indicators (solar radiation intensity I, air temperature Ta, air humidity Rh, and metal component temperature Tm) in the historical multi-source environmental parameters as independent variables, the regression coefficients of each independent variable are calculated through multiple linear regression analysis, and significant influencing factors (such as I, Rh, and Tm-Ta) with absolute regression coefficient values ​​greater than 0.05 are screened out. 3) Using significant influencing factors as input and temperature error ΔT as output, train the XGBoost model and optimize the model parameters (learning rate 0.01~0.1, tree depth 3~8, number of leaf nodes 10~30) through grid search to enable the model to accurately fit the nonlinear relationship between environmental factors and temperature error; 4) Through model interpretive analysis (such as SHAP value analysis), clarify the influence of various environmental factors on temperature error: For example, when the solar radiation intensity I ≥ 500 W / m², ΔT increases exponentially with the increase of I, and ΔT increases by an average of 0.3℃ for every 100 W / m² increase; when the temperature difference (Tm-Ta) between the metal component and the air ≥ 15℃, ΔT increases linearly with the increase of the temperature difference, and ΔT increases by an average of 0.4℃ for every 10℃ increase of the temperature difference; when the air humidity Rh ≥ 80%RH, ΔT decreases with the increase of Rh, and ΔT decreases by an average of 0.2℃ for every 10%RH increase of humidity.

[0060] It should be noted that this invention does not limit the specific algorithm for correlation analysis. Other algorithms with nonlinear fitting capabilities, such as random forests and neural networks, can also be used, as long as they can accurately uncover the patterns of environmental disturbances, and all are within the scope of protection of this invention. Here, "correlation analysis" refers to using statistical or machine learning methods to uncover the dependencies between two or more variables. The core here is to establish a quantitative mapping relationship between environmental factors and temperature errors.

[0061] Based on the environmental disturbance patterns obtained through data mining, a dynamic temperature correction coefficient generation model is constructed to achieve accurate matching between the correction coefficient and real-time multi-source environmental parameters. In this embodiment, the temperature correction coefficient is divided into a proportional correction coefficient k and an offset correction coefficient b. The proportional correction coefficient is used to correct the proportional distortion of temperature prediction caused by environmental factors, while the offset correction coefficient is used to offset system errors in a fixed direction. The generation formula of the correction coefficient is derived based on the output results of the XGBoost model. in, Let be the function that reflects the proportional distortion caused by the relationship between solar radiation intensity and metal temperature difference. This is a function representing the effect of air temperature and humidity on the offset error. , Weighting coefficients ( =0.001, =0.01, which can be calibrated using historical data.

[0062] The specific generation process is as follows: Real-time collected multi-source environmental parameters (… , , , Input the correction coefficients to generate the model. The model first calculates the significant influencing factors ( , , - Then, the pre-trained XGBoost model outputs the corresponding temperature error prediction value. Finally, according to The ratio between the predicted temperature and the predicted temperature is used to dynamically generate a proportional correction coefficient k and an offset correction coefficient b. For example, when the real-time environmental parameters are I=650W / m², Ta=25℃, Rh=60%RH, and Tm=45℃, the predicted ΔT is 1.2℃, and the predicted temperature is 42.5℃, then k=0.97 and b=-1.1℃ are generated; when the real-time environmental parameters are I=200W / m², Ta=20℃, Rh=85%RH, and Tm=22℃, ... =0.3℃, the first temperature prediction result is 31.8℃, then k=0.99, b=-0.2℃ are generated. It should be noted that the generation form of the correction coefficient can be adjusted according to the actual interference law, or a single correction coefficient (such as only offset correction or only proportional correction) can be used. This invention does not limit this, the core is to ensure that the correction coefficient can specifically counteract environmental interference.

[0063] Based on the generated temperature correction coefficient, the first temperature prediction result is corrected to obtain the final temperature detection result of the target insulator.

[0064] Preferably, a combined correction formula of "proportional correction + offset correction" is used, as follows: in, The first temperature prediction result is given, k is the proportional correction factor, and b is the offset correction factor.

[0065] In this embodiment, the correction process needs to include a rationality verification step: 1) Compare the final corrected T with the reasonable operating temperature range of the insulator (-20℃~150℃). If the final T exceeds this range, the correction coefficient is determined to be abnormal. In this case, the average correction coefficient corresponding to the same environmental parameters in history is used to correct it again. 2) Calculate the temperature change before and after the correction. ,like If the temperature exceeds 5℃ (the threshold can be adjusted according to the actual scenario), an early warning message will be output to prompt maintenance personnel to check whether the environmental parameter collection is abnormal.

[0066] For example, when At 42.5℃, k=0.97, and b=-1.1℃, =(42.5×0.97)-1.1≈40.1℃, this result is within a reasonable temperature range, and =2.4℃ < 5℃, is considered a valid correction result; if =50.3℃, k=0.85, b=-3.2℃, then =(50.3×0.85)-3.2≈40.6℃, which also meets the requirements. It should be noted that the specific form of the correction formula can be adjusted according to the type of interference. For example, when the interference is mainly linear shift, it can be simplified to... Those skilled in the art can flexibly configure the settings according to the actual situation, and all of these settings fall within the protection scope of this invention.

[0067] This embodiment utilizes historical data correlation analysis to uncover environmental interference patterns and dynamically generates targeted correction coefficients based on real-time multi-source environmental parameters, achieving accurate correction of the first temperature prediction result. This design completely overcomes the shortcomings of traditional fixed correction methods that cannot adapt to complex environments. It effectively counteracts the combined interference of multiple factors such as solar radiation, metal reflection, and air temperature and humidity, keeping the error of the final temperature detection result within ±0.3℃, improving the accuracy of the first temperature prediction result by more than 60%. Simultaneously, the correction process possesses adaptability and fault tolerance, and the reliability of the correction result is ensured through rationality verification, providing core data support for the accurate identification of zero-value insulators and the safe operation and maintenance of power systems.

[0068] Another embodiment of the present invention provides an insulator temperature detection system; for details, please refer to [link to relevant documentation]. Figure 2 , Figure 2 The diagram shown illustrates a structural block diagram of an insulator temperature detection system according to one embodiment of the present invention, comprising: Acquisition module 11 is used to acquire visible light images, infrared light images and multi-source environmental parameters of the target insulator; Registration module 12 is used to perform registration processing on the visible light image and the infrared light image to obtain a dual-light registered image; The segmentation module 13 is used to perform pixel-level segmentation on the dual-light registration image and generate an effective insulator region image based on the segmentation result; wherein, the pixel-level segmentation is designed to separate the insulator from the background based on the insulator texture features in the visible light image and the insulator temperature range features in the infrared light image. The fusion module 14 is used to extract features from the effective area image of the insulator to obtain visible light mode features and infrared light mode features, and to fuse the visible light mode features and infrared light mode features based on the multi-source environmental parameters to obtain cross-modal fusion features; Prediction module 15 is used to perform temperature prediction based on the cross-modal fusion features to obtain a first temperature prediction result; The generation module 16 is used to correct the first temperature prediction result based on multi-source environmental parameters to obtain the temperature detection result of the target insulator.

[0069] Preferably, in one embodiment of the present invention, the registration module is specifically used for: Based on the typical color and geometric features of insulators, the region of interest of the insulator in the visible light image is located. Based on the region of interest of the insulator, feature points are extracted from the visible light image and the infrared light image respectively to obtain visible light feature points and infrared light feature points; The visible light feature points and the infrared light feature points are matched using a feature matching algorithm to generate matching feature point pairs. The pixel-level mapping relationship between the visible light image and the infrared light image is determined based on the matching feature point pairs, and the image transformation is performed on the visible light image or the infrared light image based on the pixel-level mapping relationship to generate a dual-light registration image.

[0070] Preferably, in one embodiment of the present invention, the segmentation module is specifically used for: Histogram equalization is performed on the first infrared image in the dual-light registration image, and the insulator temperature range feature is extracted from the obtained second infrared image to obtain the insulator temperature range feature. Insulator texture features are extracted from the first visible light image in the dual-light registration image to obtain insulator texture features; The dual-light registration image is classified at the pixel level based on the insulator temperature range characteristics and the insulator texture characteristics. The pixel region corresponding to the target insulator is determined according to the pixel-level classification results to generate an effective insulator region image.

[0071] Preferably, in one embodiment of the present invention, the fusion module is specifically used for: The information contribution of the visible light modal features and the infrared light modal features is evaluated based on the multi-source environmental parameters, and the fusion weights of the visible light modal features and the infrared light modal features are generated according to the information contribution. The visible light modal features and the infrared light modal features are fused based on the fusion weights to obtain cross-modal fused features.

[0072] Preferably, in one embodiment of the present invention, the correction module is specifically used for: Acquire historical multi-source environmental parameters and historical temperature data of the target insulator; A correlation analysis was performed on the historical multi-source environmental parameters and the historical temperature data to obtain the interference patterns of different environmental factors on temperature detection. Based on the aforementioned interference pattern, a temperature correction coefficient corresponding to the multi-source environmental parameters is generated; The first temperature prediction result is corrected according to the temperature correction coefficient to obtain the temperature detection result of the target insulator.

[0073] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A method for detecting the temperature of an insulator, characterized in that, include: Acquire visible light images, infrared images, and multi-source environmental parameters of the target insulator; The visible light image and the infrared light image are registered to obtain a dual-light registered image; The dual-light registration image is segmented at the pixel level, and an effective insulator region image is generated based on the segmentation results; wherein, the pixel-level segmentation is designed to separate the insulator from the background based on the insulator texture features in the visible light image and the insulator temperature range features in the infrared light image. Feature extraction is performed on the image of the effective area of ​​the insulator to obtain visible light mode features and infrared light mode features. Based on the multi-source environmental parameters, the visible light mode features and the infrared light mode features are fused to obtain cross-modal fusion features. Temperature prediction is performed based on the cross-modal fusion features to obtain a first temperature prediction result; The first temperature prediction result is corrected based on multi-source environmental parameters to obtain the temperature detection result of the target insulator.

2. The insulator temperature detection method as described in claim 1, characterized in that, The registration process of the visible light image and the infrared light image to obtain a dual-light registered image includes: Based on the typical color and geometric features of insulators, the region of interest of the insulator in the visible light image is located. Based on the region of interest of the insulator, feature points are extracted from the visible light image and the infrared light image respectively to obtain visible light feature points and infrared light feature points; The visible light feature points and the infrared light feature points are matched using a feature matching algorithm to generate matching feature point pairs. The pixel-level mapping relationship between the visible light image and the infrared light image is determined based on the matching feature point pairs, and the image transformation is performed on the visible light image or the infrared light image based on the pixel-level mapping relationship to generate a dual-light registration image.

3. The insulator temperature detection method as described in claim 1, characterized in that, The step of performing pixel-level segmentation on the dual-light registration image and generating an effective insulator region image based on the segmentation results includes: Histogram equalization is performed on the first infrared image in the dual-light registration image, and the insulator temperature range feature is extracted from the obtained second infrared image to obtain the insulator temperature range feature. Insulator texture features are extracted from the first visible light image in the dual-light registration image to obtain insulator texture features; The dual-light registration image is classified at the pixel level based on the insulator temperature range characteristics and the insulator texture characteristics. The pixel region corresponding to the target insulator is determined according to the pixel-level classification results to generate an effective insulator region image.

4. The insulator temperature detection method as described in claim 1, characterized in that, The feature fusion of the visible light modal features and the infrared light modal features based on the multi-source environmental parameters to obtain cross-modal fused features includes: The information contribution of the visible light modal features and the infrared light modal features is evaluated based on the multi-source environmental parameters, and the fusion weights of the visible light modal features and the infrared light modal features are generated according to the information contribution. The visible light modal features and the infrared light modal features are fused based on the fusion weights to obtain cross-modal fused features.

5. The insulator temperature detection method as described in claim 1, characterized in that, The step of correcting the first temperature prediction result based on multi-source environmental parameters to obtain the temperature detection result of the target insulator includes: Acquire historical multi-source environmental parameters and historical temperature data of the target insulator; A correlation analysis was performed on the historical multi-source environmental parameters and the historical temperature data to obtain the interference patterns of different environmental factors on temperature detection. Based on the aforementioned interference pattern, a temperature correction coefficient corresponding to the multi-source environmental parameters is generated; The first temperature prediction result is corrected according to the temperature correction coefficient to obtain the temperature detection result of the target insulator.

6. An insulator temperature detection system, characterized in that, include: The acquisition module is used to acquire visible light images, infrared images, and multi-source environmental parameters of the target insulator; The registration module is used to register the visible light image and the infrared light image to obtain a dual-light registered image; The segmentation module is used to perform pixel-level segmentation on the dual-light registration image and generate an effective insulator region image based on the segmentation results; wherein, the pixel-level segmentation is designed to separate the insulator from the background based on the insulator texture features in the visible light image and the insulator temperature range features in the infrared light image. The fusion module is used to extract features from the effective area image of the insulator to obtain visible light mode features and infrared light mode features, and to fuse the visible light mode features and infrared light mode features based on the multi-source environmental parameters to obtain cross-modal fusion features; The prediction module is used to predict temperature based on the cross-modal fusion features to obtain a first temperature prediction result; The generation module is used to correct the first temperature prediction result based on multi-source environmental parameters to obtain the temperature detection result of the target insulator.

7. The insulator temperature detection system as described in claim 6, characterized in that, The registration module is specifically used for: Based on the typical color and geometric features of insulators, the region of interest of the insulator in the visible light image is located. Based on the region of interest of the insulator, feature points are extracted from the visible light image and the infrared light image respectively to obtain visible light feature points and infrared light feature points; The visible light feature points and the infrared light feature points are matched using a feature matching algorithm to generate matching feature point pairs. The pixel-level mapping relationship between the visible light image and the infrared light image is determined based on the matching feature point pairs, and the image transformation is performed on the visible light image or the infrared light image based on the pixel-level mapping relationship to generate a dual-light registration image.

8. The insulator temperature detection system as described in claim 6, characterized in that, The segmentation module is specifically used for: Histogram equalization is performed on the first infrared image in the dual-light registration image, and the insulator temperature range feature is extracted from the obtained second infrared image to obtain the insulator temperature range feature. Insulator texture features are extracted from the first visible light image in the dual-light registration image to obtain insulator texture features; The dual-light registration image is classified at the pixel level based on the insulator temperature range characteristics and the insulator texture characteristics. The pixel region corresponding to the target insulator is determined according to the pixel-level classification results to generate an effective insulator region image.

9. The insulator temperature detection system as described in claim 6, characterized in that, The fusion module is specifically used for: The information contribution of the visible light modal features and the infrared light modal features is evaluated based on the multi-source environmental parameters, and the fusion weights of the visible light modal features and the infrared light modal features are generated according to the information contribution. The visible light modal features and the infrared light modal features are fused based on the fusion weights to obtain cross-modal fused features.

10. The insulator temperature detection system as described in claim 6, characterized in that, The correction module is specifically used for: Acquire historical multi-source environmental parameters and historical temperature data of the target insulator; A correlation analysis was performed on the historical multi-source environmental parameters and the historical temperature data to obtain the interference patterns of different environmental factors on temperature detection. Based on the aforementioned interference pattern, a temperature correction coefficient corresponding to the multi-source environmental parameters is generated; The first temperature prediction result is corrected according to the temperature correction coefficient to obtain the temperature detection result of the target insulator.

Images