Power plant equipment voltage heating type defect detection method and system

Abstract

The application relates to the technical field of power station monitoring, and provides a power station equipment voltage heating type defect detection method and system, which comprises the following steps: acquiring an infrared image of an equipment to be detected, extracting a local image containing an equipment part through a deep learning detection model, carrying out contrast stretching processing, calculating a gray histogram and a cumulative distribution function, finding a cut-off interval in the gray histogram according to the cumulative distribution function and using a cut-off ratio, carrying out gray mapping processing on the local image according to the cut-off interval, obtaining an enhanced local image, obtaining a mask of the equipment part through an image segmentation model, converting the infrared image into an absolute temperature matrix, extracting an image area and a temperature data area corresponding to the mask in the infrared image and the absolute temperature matrix respectively, carrying out defect determination respectively, and obtaining a final voltage heating type defect determination result through decision level synthesis. The defect omission probability is greatly reduced.

Description

Technical Field

[0001] This invention belongs to the field of power plant monitoring technology, and in particular relates to a method and system for detecting voltage-heating type defects in power plant equipment. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] During the safe and stable operation of the power system, voltage-induced heating defects are particularly noteworthy due to their strong concealment, weak temperature rise characteristics, and self-accelerating development. These defects usually occur in the insulation deterioration parts inside the equipment. Early heating is not easy to detect, but it can quickly evolve into insulation breakdown or even equipment explosion.

[0004] By obtaining timely and accurate thermal status information of power equipment, it is possible to effectively identify voltage-heating defects caused by a decline in insulation performance, thereby preventing electrical faults caused by equipment overheating, which plays a key role in ensuring the reliable operation of the power grid.

[0005] Currently, voltage-induced heating defects mainly rely on manual analysis and judgment. Since the early thermal characteristics of voltage-induced heating defects are not obvious and the temperature difference is small, manual identification is not only time-consuming and labor-intensive and inefficient, but also prone to omissions and misjudgments due to insufficient experience or subjective fatigue, making it difficult to guarantee the consistency and reliability of diagnosis.

[0006] Although existing research has proposed methods for detecting defects in power equipment by fusing infrared and visible light images, which can identify infrared defects in some voltage-heated equipment such as insulators, bushings, surge arresters, current transformers, and voltage transformers, certain limitations exist:

[0007] Multi-source sensor registration challenge: It relies on the accurate registration of visible light and infrared images. However, the two imaging modalities have inherent differences in resolution, field of view and imaging principle, which makes it difficult to completely align the geometry and position of the same device in the two images, introducing positioning errors.

[0008] Limited application scenarios: Accurate detection of voltage-heated defects often requires a "pure black" environment at night, without rain or fog and with strong light sources turned off, in order to eliminate interference such as solar radiation. Under these conditions, the quality of visible light images is extremely poor, losing their value in auxiliary positioning and causing the solution to fail.

[0009] Insufficient utilization of defect features: Essentially, temperature is still used as the core criterion, failing to deeply explore the rich visual features contained in the infrared image itself (such as hot spot shape, texture anomalies, gradient changes, etc.), and the ability to identify defects that do not have significant temperature rise but have structural damage is insufficient. Summary of the Invention

[0010] To address the technical problems mentioned above, this invention provides a method and system for detecting voltage-heating defects in power plant equipment. First, a three-stage infrared equipment positioning mechanism is employed: deep learning for coarse device localization → contrast stretching and dynamic grayscale mapping fusion for local enhancement → refined generation of a mask for the equipment area. This mechanism obtains the precise contour and pixel-level position information of the equipment area. Then, a dual-channel independent criterion of vision and temperature is used to overcome the limitations of a single criterion. The vision channel effectively identifies structural defects in infrared images, while the temperature channel accurately diagnoses heat-generating defects. This complementary approach significantly reduces the possibility of missed detection of voltage-heating defects in power plant equipment.

[0011] To achieve the above objectives, the present invention adopts the following technical solution:

[0012] The first aspect of the present invention provides a method for detecting voltage-heating type defects in power plant equipment, comprising:

[0013] Deep learning device coarse localization steps: acquire infrared images of the device to be detected, and extract local images containing the device parts through a deep learning detection model;

[0014] The local enhancement steps of contrast stretching and dynamic gray-level mapping fusion are as follows: After contrast stretching the local image, the gray-level histogram and cumulative distribution function are calculated. Based on the cumulative distribution function, the cutoff interval is found in the gray-level histogram using the cutoff ratio. Based on the cutoff interval, the local image is subjected to gray-level mapping to obtain the enhanced local image.

[0015] The steps for refining the mask generation for equipment parts are as follows: Based on the enhanced local image, the mask for the equipment parts is obtained through an image segmentation model;

[0016] Decision-level defect comprehensive judgment steps: Convert the infrared image into an absolute temperature matrix. After extracting the image area and temperature data area corresponding to the mask from the infrared image and the absolute temperature matrix respectively, the final voltage heating type defect judgment result is obtained through decision-level synthesis.

[0017] Furthermore, the step of finding the cutoff interval in the grayscale histogram using the cutoff ratio includes: taking the grayscale value corresponding to the cumulative distribution function being equal to the cutoff ratio as the effective minimum grayscale value; taking the grayscale value corresponding to the cumulative distribution function being equal to 1 - the cutoff ratio as the effective maximum grayscale value; the two endpoints of the cutoff interval are the effective minimum grayscale value and the effective maximum grayscale value, respectively.

[0018] Further, the grayscale mapping processing of the local image based on the truncated interval includes: mapping the pixels within the truncated interval to the entire grayscale space, and setting the pixel values ​​outside the truncated interval as the endpoint values ​​of the entire grayscale space to obtain an enhanced local image; the formula for mapping the pixels within the truncated interval to the entire grayscale space is: G'=clamp(0,255,(G_original-G_min_eff)×(255 / (G_max_eff-G_min_eff))); where G' is the mapped pixel value, G_original is the pixel within the truncated interval, G_min_eff is the effective minimum grayscale value, and G_max_eff is the effective maximum grayscale value.

[0019] Furthermore, the truncation ratio is determined based on the gray-level mode and gray-level variance in the gray-level histogram, and the truncation ratio is... Where p0 is the basic cutoff ratio, The variance is the reference variance, and a and b are the weighting coefficients. This represents the grayscale variance of a local image after contrast stretching. G represents the mode deviation of gray levels in a local image after contrast stretching. min G represents the minimum grayscale value of a local image after contrast stretching. max This represents the maximum grayscale value of the local image after contrast stretching.

[0020] Furthermore, for enhancing local images, after segmentation using an image segmentation model, the area thresholding method is used to remove noise regions from the segmentation result image, and morphological closing operation and Gaussian filtering are used for smoothing to obtain a mask for the device area.

[0021] Furthermore, a lightweight convolutional neural network is used to determine defects in the image region, and channel attention and feature pyramid structure are embedded in the lightweight convolutional neural network.

[0022] Furthermore, the steps for defect determination in the temperature data area include: for the temperature data area, statistically analyzing the highest temperature, absolute temperature difference, and relative temperature difference, and determining defects based on the statistical information.

[0023] A second aspect of the present invention provides a voltage-heating type defect detection system for power plant equipment, comprising:

[0024] Deep learning device coarse localization steps: acquire infrared images of the device to be detected, and extract local images containing the device parts through a deep learning detection model;

[0025] The local enhancement steps of contrast stretching and dynamic gray-level mapping fusion are as follows: After contrast stretching the local image, the gray-level histogram and cumulative distribution function are calculated. Based on the cumulative distribution function, the cutoff interval is found in the gray-level histogram using the cutoff ratio. Based on the cutoff interval, the local image is subjected to gray-level mapping to obtain the enhanced local image.

[0026] The steps for refining the mask generation for equipment parts are as follows: Based on the enhanced local image, the mask for the equipment parts is obtained through an image segmentation model;

[0027] Decision-level defect comprehensive judgment steps: Convert the infrared image into an absolute temperature matrix. After extracting the image area and temperature data area corresponding to the mask from the infrared image and the absolute temperature matrix respectively, the final voltage heating type defect judgment result is obtained through decision-level synthesis.

[0028] Furthermore, the step of finding the cutoff interval in the grayscale histogram using the cutoff ratio includes: taking the grayscale value corresponding to the cumulative distribution function being equal to the cutoff ratio as the effective minimum grayscale value; taking the grayscale value corresponding to the cumulative distribution function being equal to 1 - the cutoff ratio as the effective maximum grayscale value; the two endpoints of the cutoff interval are the effective minimum grayscale value and the effective maximum grayscale value, respectively.

[0029] Further, the grayscale mapping processing of the local image based on the truncated interval includes: mapping the pixels within the truncated interval to the entire grayscale space, and setting the pixel values ​​outside the truncated interval as the endpoint values ​​of the entire grayscale space to obtain an enhanced local image; the formula for mapping the pixels within the truncated interval to the entire grayscale space is: G'=clamp(0,255,(G_original-G_min_eff)×(255 / (G_max_eff-G_min_eff))); where G' is the mapped pixel value, G_original is the pixel within the truncated interval, G_min_eff is the effective minimum grayscale value, and G_max_eff is the effective maximum grayscale value.

[0030] Furthermore, the truncation ratio is determined based on the gray-level mode and gray-level variance in the gray-level histogram, and the truncation ratio is... Where p0 is the basic cutoff ratio, The variance is the reference variance, and a and b are the weighting coefficients. This represents the grayscale variance of a local image after contrast stretching. G represents the mode deviation of gray levels in a local image after contrast stretching. min G represents the minimum grayscale value of a local image after contrast stretching. max This represents the maximum grayscale value of the local image after contrast stretching.

[0031] Furthermore, for enhancing local images, after segmentation using an image segmentation model, the area thresholding method is used to remove noise regions from the segmentation result image, and morphological closing operation and Gaussian filtering are used for smoothing to obtain a mask for the device area.

[0032] Furthermore, a lightweight convolutional neural network is used to determine defects in the image region, and channel attention and feature pyramid structure are embedded in the lightweight convolutional neural network.

[0033] Furthermore, the steps for defect determination in the temperature data area include: for the temperature data area, statistically analyzing the highest temperature, absolute temperature difference, and relative temperature difference, and determining defects based on the statistical information.

[0034] A third aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described method for detecting voltage-heating defects in power plant equipment.

[0035] A fourth aspect of the present invention provides a computer device including a computer-readable storage medium, a processor, and a computer program stored on the computer-readable storage medium and executable on the processor, wherein the processor executes the program to implement the steps of the above-described method for detecting voltage-heating defects in power plant equipment.

[0036] Compared with the prior art, the beneficial effects of the present invention are:

[0037] This invention innovatively proposes a method for detecting voltage-heating defects in power plant equipment. First, it employs a three-segment infrared equipment localization mechanism: deep learning for coarse equipment localization, contrast stretching and dynamic grayscale mapping fusion for local enhancement, and fine generation of equipment part masks. This mechanism obtains the precise contour and pixel-level position information of the equipment parts. Then, it adopts a dual-channel independent criterion of vision and temperature to overcome the limitations of a single criterion. The vision channel can effectively identify structural defects in infrared images, while the temperature channel can accurately diagnose heat-generating defects. The complementarity of the two greatly reduces the possibility of missing voltage-heating defects in power plant equipment.

[0038] This invention innovatively proposes a local enhancement algorithm that fuses contrast stretching and dynamic grayscale mapping. First, it improves local contrast and suppresses noise by limiting the contrast through adaptive histogram equalization. Then, it adopts a dynamic range stretching method based on percentile truncation to automatically eliminate extreme noise points and linearly maps the effective grayscale range of the image to the full range [0,255], maximizing the use of the grayscale space and highlighting the detailed features of the device under test. It is suitable for multiple application scenarios.

[0039] This invention innovatively proposes a decision-level defect comprehensive judgment method, which adopts a dual-channel defect judgment with independent vision and temperature. The vision channel uses a lightweight convolutional neural network to focus on identifying structural defects, while the temperature channel performs heat-related defect diagnosis on the equipment by calculating quantitative indicators such as temperature difference. Finally, a hierarchical decision is used to integrate the results of the two channels to obtain a unique, reliable and interpretable final conclusion, which combines the high sensitivity of deep learning with the high reliability of standard rules. Attached Figure Description

[0040] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0041] Figure 1 This is a flowchart of a voltage-heating type defect detection method for power plant equipment according to Embodiment 1 of the present invention;

[0042] Figure 2 This is a flowchart of the local infrared enhancement process in Embodiment 1 of the present invention;

[0043] Figure 3 This is a comparison of infrared images before and after local enhancement by contrast stretching and dynamic grayscale mapping fusion according to Embodiment 1 of the present invention.

[0044] Figure 4 This is a structural diagram of the image segmentation model of Embodiment 1 of the present invention;

[0045] Figure 5 This is a structural diagram of the defect identification model according to Embodiment 1 of the present invention;

[0046] Figure 6 This is a schematic diagram of the structure of a computer device according to Embodiment 4 of the present invention. Detailed Implementation

[0047] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0048] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0049] Example 1

[0050] This embodiment provides a method for detecting voltage-heating type defects in power plant equipment.

[0051] This embodiment provides a method for detecting voltage-heating type defects in power plant equipment, which is applicable to substations.

[0052] This embodiment provides a method for detecting voltage-heating type defects in power plant equipment. It can keenly capture the characteristics of low contrast and small temperature difference thermal anomalies, and realize the intelligent diagnosis of voltage-heating type defects automatically, quickly and objectively. This provides a key basis for timely early warning and handling, improves the efficiency and accuracy of power plant equipment status management, and provides technical support for the safe and stable operation of the power system.

[0053] To address the practical application needs of voltage-heating type defect detection in power plants, and the shortcomings of existing detection methods in terms of anti-interference capability and defect judgment, this embodiment provides a method for detecting voltage-heating type defects in power plant equipment, aiming to solve the following two key technical problems:

[0054] First, to overcome the problem of complex background heat sources and small temperature differences in the target area in infrared images leading to large device positioning deviations, this embodiment introduces a deep learning target detection network to achieve preliminary identification and positioning of target devices in infrared images. Combining local enhancement technology of contrast stretching and dynamic grayscale mapping fusion in the target area of ​​infrared images with deep learning segmentation methods, it achieves fine extraction of the heat-generating area of ​​the device and effectively suppresses high background temperature and environmental noise interference.

[0055] Second, the method for judging voltage-heating defects is optimized. Based on the statistical analysis of infrared temperature matrix, a dedicated infrared feature library for voltage-heating defects is constructed, and an infrared defect judgment method based on image features is added. This achieves dual-channel collaborative judgment of "temperature parameters + visual features" and avoids the misjudgment problem of single temperature analysis.

[0056] This embodiment provides a method for detecting voltage-heating type defects in power plant equipment. It achieves the analysis and judgment of voltage-heating type defects through two parts: multi-level precise positioning of the equipment location using infrared technology and dual-channel defect judgment. Figure 1 As shown, the specific steps include the following:

[0057] Step 1: Multi-level precise positioning of infrared equipment components.

[0058] By employing a three-tiered processing flow of deep learning for coarse localization of infrared devices, image enhancement, and refined generation of device part masks, efficient and accurate localization of infrared device parts can be achieved.

[0059] Step 101: Coarse localization of deep learning devices based on deep learning detection models.

[0060] (1) Dataset construction.

[0061] First, an infrared equipment component dataset is constructed. This dataset contains images of infrared equipment of different types and under different operating conditions, covering various states such as normal equipment temperature and the presence of thermal defects. At the same time, the bounding box information of key components such as ceramic sleeves and electromagnetic units in the infrared equipment images is annotated to ensure the diversity and accuracy of the dataset.

[0062] (2) Training of deep learning detection model.

[0063] The YOLOv8 model was selected as the coarse localization model for infrared image target deep learning devices (i.e., deep learning detection model). The deep learning detection model was trained using the constructed infrared device part dataset. The adaptive moment estimation (Adam) optimization algorithm was adopted, and the deep learning detection model was converged through iterative training. Finally, a deep learning detection model that meets the requirements of coarse localization of infrared device parts was obtained.

[0064] (3) Coarse localization execution process of deep learning device.

[0065] The infrared image of the device to be located (i.e., the infrared image of the device to be detected) is input into a trained deep learning detection model. The deep learning detection model first extracts features from the infrared device image by using convolutional layers, pooling layers, and feature fusion layers to gradually extract low-level features (such as edges and textures) and high-level features (such as device component outlines) from the infrared device image. The region generation network (RPN) or target prediction head of the deep learning detection model is used to analyze the extracted feature map, predict candidate regions in the infrared device image that may contain device parts, and calculate the confidence score (representing the probability that the region is a device part) and bounding box coordinates for each candidate region.

[0066] Set a confidence threshold and filter out candidate regions with confidence scores higher than the threshold. These candidate regions are the approximate locations of the device parts in the infrared device image, completing the deep learning device coarse localization process. At this point, the obtained deep learning device coarse localization result can determine the approximate range of the device parts, but the bounding box may include some background areas and cannot accurately reflect the specific outline of the device parts.

[0067] Step 102: Image contrast stretching based on local infrared enhancement technology.

[0068] (1) Boundary expansion.

[0069] Based on the bounding box coordinates of the device parts obtained from the coarse localization by the deep learning device, a local image region containing the device parts is extracted from the original infrared device image. To avoid missing the detailed information of the device parts edges during the subsequent enhancement process, the bounding box is expanded to a certain pixel range in all directions when extracting the local region (the number of expanded pixels is set to 5-20 pixels, which is automatically adjusted according to the size of the device parts and the image resolution), to obtain the local image to be enhanced.

[0070] (2) Local infrared enhancement processing based on CLAHE algorithm and dynamic range adjustment.

[0071] To improve the accuracy of subsequent image segmentation, this embodiment adopts an advanced two-stage local image enhancement process: first, the CLAHE algorithm is used to optimize local contrast and suppress noise; then, the utilization rate of gray space is maximized by calculating the dynamic range based on statistics, thereby providing the optimal input image for the device part mask refinement generation model.

[0072] Specific procedures, such as Figure 2 As shown, it includes:

[0073] (A) The Adaptive Histogram Equalization (CLAHE) algorithm, combined with the grayscale characteristics of infrared images, is used to perform contrast stretching on the local image to be enhanced. Compared with the traditional global histogram equalization algorithm, the CLAHE algorithm can divide the local image into multiple sub-blocks (N×N sub-modules), perform histogram equalization on each sub-block separately, and at the same time, by limiting the contrast threshold, avoid the problem of over-enhancement of the image or noise amplification caused by the concentration of grayscale values ​​in local areas.

[0074] (B) The grayscale dynamic range of the image (I_clahe) after CLAHE processing may still not fully cover the [0,255] interval. Therefore, a dynamic range calculation and stretching step based on image content is used:

[0075] First, calculate the gray-level statistical histogram, and calculate the gray-level histogram and cumulative distribution function (CDF) for the I_clahe image.

[0076] Secondly, a cutoff ratio p is set;

[0077] Then, based on the truncation ratio p, the effective minimum gray value (G_min_eff) is calculated, and the gray value corresponding to the CDF value is found when it equals p%. This means that if p% of the pixel gray values ​​are lower than this value, they will be regarded as extremely dark noise and ignored.

[0078] Then, based on the truncation ratio p, calculate the effective maximum gray value (G_max_eff): Find the gray value corresponding to the CDF value equal to (100 - p)%. This means that p% of the pixel gray values are higher than this value, and they will be regarded as extremely bright noise and ignored.

[0079] Finally, perform linear stretching: Linearly map the pixel G_original with gray value in the truncation interval [G_min_eff, G_max_eff] in the I_clahe image to the entire 8-bit gray space [0, 255]. The transformation formula is:

[0080] G' = clamp(0, 255, (G_original - G_min_eff) × (255 / (G_max_eff - G_min_eff)));

[0081] For pixels with G_original < G_min_eff, directly set them to 0 (pure black); for pixels with G_original > G_max_eff, directly set them to 255 (pure white).

[0082] As Figure 3 shown, through the above processing, the gray contrast between the device part and the background can be significantly improved, making the details such as the edges and textures of the device part clearer, laying a foundation for the subsequent refined generation of the device part mask.

[0083] As an implementation, the truncation ratio p is set by technical experts.

[0084] As another implementation, the truncation ratio p is calculated through the following steps:

[0085] Let the gray variance of the I_clahe image be , m be the mode of the gray value of the I_clahe image (the gray value with the highest occurrence frequency), G min be the minimum gray value of the I_clahe image, G max be the maximum gray value of the I_clahe image;

[0086] First, calculate the mode deviation d m , used to measure the deviation degree of the mode from the center of the gray range: d m = |m - (G min + G max ) / 2|;

[0087] Then, comprehensively consider the variance and the mode deviation, and calculate the truncation ratio: ; where p0 is the basic truncation ratio (determined through experiments, such as 1%); The reference variance is obtained statistically from images of the same type with ideal imaging effects, representing the variance with a moderate gray-scale distribution; a and b are weighting coefficients (which need to be adjusted experimentally to control the influence of variance and mode deviation on the cutoff ratio, and are generally taken as values ​​between 0.2 and 0.8).

[0088] The difference between the current image's grayscale dispersion and a reference dispersion is measured. Much larger This indicates that the grayscale distribution is too dispersed (containing many extremely dark or extremely bright pixels). In this case, the truncation ratio will be increased to ignore more noise or abnormal pixels; if much smaller This indicates that the grayscale distribution is too concentrated. This part will also increase the truncation ratio, expanding the relative range of the effective grayscale range by ignoring extreme pixels.

[0089] mode deviation term This measures the relative degree to which the mode deviates from the center of the grayscale range. The further the mode deviates from the center, the more the concentrated area of ​​the grayscale distribution deviates from the midtone. This part will adjust the cutoff ratio to make the effective grayscale range more in line with the grayscale distribution of the target object (for example, if the target is in a darker area, by adjusting the cutoff ratio, the effective darker area can be stretched more fully).

[0090] Step 103: Refined generation of device part masks based on image segmentation technology.

[0091] (1) Selection and optimization of segmentation model.

[0092] U-Net++ (an enhanced version of U-Net, where U-Net is a deep learning model for image segmentation) was chosen as the image segmentation model for refining the mask of infrared device parts. Considering that infrared device part segmentation needs to balance accuracy and efficiency, the image segmentation model was optimized as follows: First, a lightweight network, EfficientNet-Lite (a lightweight visual model), was introduced into the encoder part of the image segmentation model to reduce the number of parameters and improve the segmentation speed. Second, the number and dimension of skip connections were increased in the decoder part of the image segmentation model to enhance the ability to recover detailed features of the device parts. Third, a combined loss function of Dice loss (also known as F1 loss) and cross-entropy loss was introduced into the loss function of the image segmentation model to solve the class imbalance problem that may exist in the segmentation of infrared device parts and improve the segmentation accuracy of the image segmentation model for device parts.

[0093] The optimized image segmentation model was trained using the constructed infrared device part dataset (with annotation information expanded to pixel-level masks of device parts). The training process adopted an optimization strategy similar to that of the deep learning detection model to ensure that the model could accurately segment pixel-level regions of device parts in infrared images.

[0094] like Figure 4 As shown, the optimized image segmentation model includes an encoder and a decoder; the encoder uses the device infrared spectrum preprocessed by the lightweight network EfficientNet-Lite; the decoder includes sequentially connected upsampling and convolutional parts, as well as feature fusion and dense skip connections.

[0095] (2) Execution process of fine-grained generation of mask for equipment parts.

[0096] The local image after local infrared enhancement (enhanced local image) is input into the trained image segmentation model. The model extracts features from the enhanced local image through the encoder to obtain feature maps at different scales. Then, the feature maps are upsampled and feature fused through the decoder to gradually restore the spatial resolution of the image. Finally, the segmentation result map with the same size as the local image is output (in the segmentation result map, the device area is marked with a specific pixel value, and the background area is marked with other pixel values).

[0097] The segmentation results are post-processed by using the area thresholding method to remove small noise areas from the segmentation results, and morphological closing operation and Gaussian filtering to smooth the edges of the device parts, resulting in the final refined generation result of the device parts mask.

[0098] Based on the positional correspondence between the local image and the original infrared device image, the refined mask generation result of the device part is mapped back to the original image to determine the accurate pixel-level position information of the device part in the original image, including the outline of the device part, the coordinates of key detail areas, etc., to obtain the precise pixel-level mask of the target device part, thus completing the entire multi-level infrared device part precise positioning process.

[0099] Step 2: A dual-channel infrared defect determination method based on visual and temperature matrices.

[0100] This embodiment proposes an independent dual-channel criterion mechanism. This mechanism assesses the device status through two independent and parallel paths: visual morphology analysis and temperature distribution analysis. Finally, it combines the two results based on clear decision rules to form a final judgment.

[0101] Step 201: Data preprocessing and ROI (Region of Interest) extraction.

[0102] (1) Input data: raw infrared image and its corresponding absolute temperature matrix; precise pixel-level mask of the target device.

[0103] (2) Target region extraction: Using the above mask, extract the image region and temperature data region containing only the target device part from the original infrared image and absolute temperature matrix.

[0104] Step 202: Dual-channel analysis and judgment.

[0105] (1) Channel A: Visual morphological defect determination based on deep learning.

[0106] (A) Dataset construction.

[0107] To address the reality that there are far more "normal" samples than "defective" samples, enhancement techniques are used to expand defective samples when constructing the training set: in addition to conventional rotation and flipping, proprietary enhancement techniques such as simulating infrared imaging characteristics, applying local contrast-limited enhancement, adding Gaussian thermal noise, and simulating thermal diffusion effects are applied to greatly enrich the diversity of training data and improve the generalization ability of the model.

[0108] (B) Defect identification model training.

[0109] Lightweight Network Selection and Improvement: To meet the stringent requirements of low power consumption and high speed for on-site deployment, the extremely lightweight SqueezeNet (lightweight convolutional neural network) was selected as the defect recognition model, and specific optimizations were made for infrared defect features, such as... Figure 5 As shown:

[0110] Embedded attention mechanism: A lightweight channel attention module (such as SE Block) is embedded in the Fire module, enabling the lightweight network to adaptively focus on key feature channels related to defects, suppress background interference, and effectively improve the ability to perceive subtle defects in low-contrast infrared images.

[0111] Multi-scale feature fusion: A feature pyramid structure is introduced in the back end of the lightweight network to fuse feature maps from different depths, enabling the defect recognition model to simultaneously perceive global semantic information (determine the type of defect) and local detail information (locate the edge of the defect), which is especially suitable for defect targets of different sizes;

[0112] Targeted loss function design: To completely solve the class imbalance problem, the standard cross-entropy loss is abandoned, and Focal Loss is adopted as the loss function for model training. Focal Loss reduces the weight of "easy-to-classify samples", making the defect recognition model training process more focused on "difficult-to-classify samples" (such as minor cracks and slight dirt), thereby significantly improving the detection rate of defects, especially small target defects.

[0113] Defect identification models include:

[0114] Input layer: Receives infrared defect images, then performs initial convolutional layer operations to perform preliminary feature extraction on the input image, and then performs downsampling on the feature map through a max pooling layer to reduce the number of parameters and computational load while retaining the main features;

[0115] Backbone Network: In the backbone network, the combination operation of Fire module and SE Block (squeeze-excitation block) is repeated (including Fire 2 and SE Block, Fire 3 and SE Block, Fire 4 and SE Block, Fire 5 and SE Block, ..., Fire N and SE Block). Fire module is a lightweight convolutional module that can extract effective features while reducing the amount of computation. Repeated combinations can gradually extract more abstract and higher-level image features.

[0116] Feature fusion stage: A 1×1 convolution operation is performed on the features output by Fire 5 and SE Block to obtain the first feature; a 1×1 convolution operation and upsampling are performed on the features output by Fire N and SE Block to obtain the second feature; the first and second features are fused and then upsampled through 2×2 convolution kernels to obtain the third feature; the features output by Fire 4 and SE Block are performed with a 1×1 convolution operation and then fused with the third feature to obtain the final feature. Through these operations, features at different levels and scales are integrated to obtain a more comprehensive and effective feature representation, facilitating subsequent defect identification.

[0117] (C) Output: The output of this channel is the preliminary visual morphology judgment result Result_Visual, which includes defect category labels (e.g., insulator damage), confidence probability, and optional class activation heatmap, providing rich and reliable basis for subsequent comprehensive decision-making.

[0118] (2) Channel B: Determination of temperature distribution based on statistics.

[0119] Infrared thermal defects of equipment are judged based on surface temperature judgment method and relative temperature difference judgment method.

[0120] The following three parameters are calculated for the temperature matrix Temp_Matrix_Mask in the mask region:

[0121] Maximum temperature of equipment section (T_max): The highest surface temperature of the object being measured;

[0122] Absolute temperature difference (T_diff): The temperature difference between the surfaces of the objects being measured;

[0123] Relative temperature difference (ΔT): ,in, For hot spot temperature, This is the temperature within the normal temperature range. The ambient temperature of the area being measured.

[0124] Based on the statistical information of local temperature, determine the defect type and defect level.

[0125] Step 203: Decision-level synthesis and final output.

[0126]

[0127] This embodiment proposes a three-stage infrared device localization mechanism that progresses from coarse to fine, innovatively combining target detection with image segmentation techniques. First, a high-speed deep learning coarse localization of the device is performed using an improved YOLOv8 model (introducing an attention mechanism and a complete intersection-union loss function CIoU) to determine the approximate device region. Then, a two-stage enhancement algorithm combining CLAHE and dynamic range adjustment is used to optimize the contrast of local regions, laying the foundation for segmentation. Finally, a segmentation model based on lightweight U-Net++ (using EfficientNet-Lite as the encoder and a combined loss function) is used to achieve pixel-level fine-grained generation of device region masks. Through a three-stage workflow—"coarse device localization using deep learning (YOLOv8) → local enhancement through contrast stretching and dynamic grayscale mapping (CLAHE + dynamic stretching) → refined generation of device region masks (U-Net++)"—precise contours and pixel-level positional information of the device region can be obtained. Furthermore, the coarse device localization process employs a lightweight deep learning detection model, enabling rapid initial localization of the device region and avoiding the time-consuming complex processing of the entire image. Simultaneously, the local enhancement and segmentation through contrast stretching and dynamic grayscale mapping are performed only on local areas of the device region, further improving processing efficiency. This approach significantly enhances processing efficiency while maintaining high accuracy.

[0128] This embodiment proposes a local adaptive enhancement algorithm for infrared image characteristics, overcoming the limitations of traditional global enhancement. First, it enhances local contrast and suppresses noise by using contrast-limited adaptive histogram equalization (CLAHE). Then, it employs a percentile-truncation-based dynamic range stretching method to automatically eliminate extreme noise points, linearly mapping the effective grayscale range of the image to the full range [0, 255], maximizing the use of the grayscale space and highlighting the detailed features of the device.

[0129] This embodiment proposes a defect identification and fusion decision algorithm with independent visual and temperature channels, pioneering a "two-pronged" criterion approach. The visual channel employs a lightweight SqueezeNet network improved with attention mechanisms and multi-scale feature fusion, and applies Focal Loss and infrared characteristic data enhancement techniques to focus on identifying structural defects. The temperature channel, based on industry standards, diagnoses heat-related defects by calculating quantitative indicators such as temperature difference. Finally, a hierarchical decision integrates the results from both channels to arrive at a unique, reliable, and interpretable final conclusion, combining the high sensitivity of deep learning with the high reliability of standard rules. The use of independent visual-temperature dual-channel criteria overcomes the limitations of single criteria. The visual channel effectively identifies structural defects in infrared images, while the temperature channel accurately diagnoses heat-related defects; their complementarity greatly reduces the possibility of missed detections.

[0130] Example 2

[0131] This embodiment provides a voltage-heating type defect detection system for power plant equipment, comprising:

[0132] The deep learning device coarse localization module is configured to: acquire infrared images of the device to be detected, and extract local images containing the device parts through a deep learning detection model;

[0133] The local enhancement module that fuses contrast stretching and dynamic grayscale mapping is configured to: after contrast stretching the local image, calculate the grayscale histogram and cumulative distribution function, and find the cutoff interval in the grayscale histogram based on the cumulative distribution function and the cutoff ratio, and perform grayscale mapping on the local image based on the cutoff interval to obtain the enhanced local image.

[0134] The device part mask refinement generation module is configured to: obtain the device part mask based on the enhanced local image and through the image segmentation model;

[0135] The decision-level defect comprehensive judgment module is configured to: convert the infrared image into an absolute temperature matrix, extract the image area and temperature data area corresponding to the mask from the infrared image and the absolute temperature matrix respectively, and then obtain the final voltage heating type defect judgment result through decision-level synthesis.

[0136] Furthermore, the step of finding the cutoff interval in the grayscale histogram using the cutoff ratio includes: taking the grayscale value corresponding to the cumulative distribution function being equal to the cutoff ratio as the effective minimum grayscale value; taking the grayscale value corresponding to the cumulative distribution function being equal to 1 - the cutoff ratio as the effective maximum grayscale value; the two endpoints of the cutoff interval are the effective minimum grayscale value and the effective maximum grayscale value, respectively.

[0137] Further, the grayscale mapping processing of the local image based on the truncated interval includes: mapping the pixels within the truncated interval to the entire grayscale space, and setting the pixel values ​​outside the truncated interval as the endpoint values ​​of the entire grayscale space to obtain an enhanced local image; the formula for mapping the pixels within the truncated interval to the entire grayscale space is: G'=clamp(0,255,(G_original-G_min_eff)×(255 / (G_max_eff-G_min_eff))); where G' is the mapped pixel value, G_original is the pixel within the truncated interval, G_min_eff is the effective minimum grayscale value, and G_max_eff is the effective maximum grayscale value.

[0138] Furthermore, the truncation ratio is determined based on the gray-level mode and gray-level variance in the gray-level histogram, and the truncation ratio is... Where p0 is the basic cutoff ratio, The variance is the reference variance, and a and b are the weighting coefficients. This represents the grayscale variance of a local image after contrast stretching. G represents the mode deviation of gray levels in a local image after contrast stretching. min G represents the minimum grayscale value of a local image after contrast stretching. max This represents the maximum grayscale value of the local image after contrast stretching.

[0139] Furthermore, for enhancing local images, after segmentation using an image segmentation model, the area thresholding method is used to remove noise regions from the segmentation result image, and morphological closing operation and Gaussian filtering are used for smoothing to obtain a mask for the device area.

[0140] Furthermore, a lightweight convolutional neural network is used to determine defects in the image region, and channel attention and feature pyramid structure are embedded in the lightweight convolutional neural network.

[0141] Furthermore, the steps for defect determination in the temperature data area include: for the temperature data area, statistically analyzing the highest temperature, absolute temperature difference, and relative temperature difference, and determining defects based on the statistical information.

[0142] It should be noted that each module in this embodiment corresponds one-to-one with each step in Embodiment 1, and their specific implementation processes are the same, so they will not be repeated here.

[0143] Example 3

[0144] This embodiment provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps in the voltage heating type defect detection method for power plant equipment as described in Embodiment 1 above.

[0145] Example 4

[0146] This embodiment provides a computer device, such as... Figure 6 As shown, the system includes a computer-readable storage medium 1003, a processor 1001, a communication interface 1002, and a computer program stored on the computer-readable storage medium 1003 and executable on the processor 1001. The processor 1001, communication interface 1002, and computer-readable storage medium 1003 can be connected via a bus or other means. The communication interface 1002 is used to receive and send data. When the processor 1001 executes the program, it implements the steps in the voltage heating type defect detection method for power plant equipment as described in Embodiment 1 above.

[0147] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for detecting voltage-heating type defects in power plant equipment, characterized in that, include: Deep learning device coarse localization steps: acquire infrared images of the device to be detected, and extract local images containing the device parts through a deep learning detection model; The local enhancement steps of contrast stretching and dynamic gray-level mapping fusion are as follows: After contrast stretching the local image, the gray-level histogram and cumulative distribution function are calculated. Based on the cumulative distribution function, the cutoff interval is found in the gray-level histogram using the cutoff ratio. Based on the cutoff interval, the local image is subjected to gray-level mapping to obtain the enhanced local image. The steps for refining the mask generation for equipment parts are as follows: Based on the enhanced local image, the mask for the equipment parts is obtained through an image segmentation model; Decision-level defect comprehensive judgment steps: Convert the infrared image into an absolute temperature matrix. After extracting the image area and temperature data area corresponding to the mask from the infrared image and the absolute temperature matrix respectively, the final voltage heating type defect judgment result is obtained through decision-level synthesis.

2. The method for detecting voltage-heating type defects in power plant equipment as described in claim 1, characterized in that, The step of finding the cutoff interval in the grayscale histogram using the cutoff ratio includes: taking the grayscale value corresponding to the cumulative distribution function being equal to the cutoff ratio as the effective minimum grayscale value; taking the grayscale value corresponding to the cumulative distribution function being equal to 1 minus the cutoff ratio as the effective maximum grayscale value; the two endpoints of the cutoff interval are the effective minimum grayscale value and the effective maximum grayscale value, respectively.

3. The method for detecting voltage-heating type defects in power plant equipment as described in claim 1, characterized in that, The grayscale mapping process of the local image based on the truncated interval includes: mapping the pixels within the truncated interval to the entire grayscale space, and setting the pixel values ​​outside the truncated interval as the endpoint values ​​of the entire grayscale space to obtain an enhanced local image; the formula for mapping the pixels within the truncated interval to the entire grayscale space is: G'=clamp(0,255,(G_original-G_min_eff)×(255 / (G_max_eff-G_min_eff))); where G' is the mapped pixel value, G_original is the pixel within the truncated interval, G_min_eff is the effective minimum grayscale value, and G_max_eff is the effective maximum grayscale value.

4. The method for detecting voltage-heating type defects in power plant equipment as described in claim 1, characterized in that, The truncation ratio is determined based on the gray-level mode and gray-level variance in the gray-level histogram. The truncation ratio is... Where p0 is the basic cutoff ratio, The variance is the reference variance, and a and b are the weighting coefficients. This represents the grayscale variance of a local image after contrast stretching. G represents the mode deviation of gray levels in a local image after contrast stretching. min G represents the minimum grayscale value of a local image after contrast stretching. max This represents the maximum grayscale value of the local image after contrast stretching.

5. The method for detecting voltage-heating type defects in power plant equipment as described in claim 1, characterized in that, To enhance local images, after segmentation using an image segmentation model, the area thresholding method is used to remove noise regions from the segmentation results, and morphological closing operation and Gaussian filtering are used for smoothing to obtain a mask for the device area.

6. The method for detecting voltage-heating type defects in power plant equipment as described in claim 1, characterized in that, For image regions, a lightweight convolutional neural network is used for defect detection, and channel attention and feature pyramid structures are embedded in the lightweight convolutional neural network.

7. The method for detecting voltage-heating type defects in power plant equipment as described in claim 1, characterized in that, The steps for defect determination in temperature data areas include: for the temperature data area, statistically analyzing the highest temperature, absolute temperature difference, and relative temperature difference, and determining defects based on the statistical information.

8. A voltage-heating type defect detection system for power plant equipment, characterized in that, include: The deep learning device coarse localization module is configured to: acquire infrared images of the device to be detected, and extract local images containing the device parts through a deep learning detection model; The local enhancement module that fuses contrast stretching and dynamic grayscale mapping is configured to: after contrast stretching the local image, calculate the grayscale histogram and cumulative distribution function, and find the cutoff interval in the grayscale histogram based on the cumulative distribution function and the cutoff ratio, and perform grayscale mapping on the local image based on the cutoff interval to obtain the enhanced local image. The device part mask refinement generation module is configured to: obtain the device part mask based on the enhanced local image and through the image segmentation model; The decision-level defect comprehensive judgment module is configured to: convert the infrared image into an absolute temperature matrix, extract the image area and temperature data area corresponding to the mask from the infrared image and the absolute temperature matrix respectively, and then obtain the final voltage heating type defect judgment result through decision-level synthesis.

9. The power plant equipment voltage heating type defect detection system as described in claim 8, characterized in that, The step of finding the cutoff interval in the grayscale histogram using the cutoff ratio includes: taking the grayscale value corresponding to the cumulative distribution function being equal to the cutoff ratio as the effective minimum grayscale value; taking the grayscale value corresponding to the cumulative distribution function being equal to 1 minus the cutoff ratio as the effective maximum grayscale value; the two endpoints of the cutoff interval are the effective minimum grayscale value and the effective maximum grayscale value, respectively.

10. The power plant equipment voltage heating type defect detection system as described in claim 8, characterized in that, The grayscale mapping process of the local image based on the truncated interval includes: mapping the pixels within the truncated interval to the entire grayscale space, and setting the pixel values ​​outside the truncated interval as the endpoint values ​​of the entire grayscale space to obtain an enhanced local image; the formula for mapping the pixels within the truncated interval to the entire grayscale space is: G'=clamp(0,255,(G_original-G_min_eff)×(255 / (G_max_eff-G_min_eff))); where G' is the mapped pixel value, G_original is the pixel within the truncated interval, G_min_eff is the effective minimum grayscale value, and G_max_eff is the effective maximum grayscale value.

11. The power plant equipment voltage heating type defect detection system as described in claim 8, characterized in that, The truncation ratio is determined based on the gray-level mode and gray-level variance in the gray-level histogram. The truncation ratio is... Where p0 is the basic cutoff ratio, The variance is the reference variance, and a and b are the weighting coefficients. This represents the grayscale variance of a local image after contrast stretching. G represents the mode deviation of gray levels in a local image after contrast stretching. min G represents the minimum grayscale value of a local image after contrast stretching. max This represents the maximum grayscale value of the local image after contrast stretching.

12. The power plant equipment voltage heating type defect detection system as described in claim 8, characterized in that, To enhance local images, after segmentation using an image segmentation model, the area thresholding method is used to remove noise regions from the segmentation results, and morphological closing operation and Gaussian filtering are used for smoothing to obtain a mask for the device area.

13. The power plant equipment voltage heating type defect detection system as described in claim 8, characterized in that, For image regions, a lightweight convolutional neural network is used for defect detection, and channel attention and feature pyramid structures are embedded in the lightweight convolutional neural network.

14. The power plant equipment voltage heating type defect detection system as described in claim 8, characterized in that, The steps for defect determination in temperature data areas include: for the temperature data area, statistically analyzing the highest temperature, absolute temperature difference, and relative temperature difference, and determining defects based on the statistical information.

15. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps in the method for detecting voltage-heating defects in power plant equipment as described in any one of claims 1-7.

16. A computer device comprising a computer-readable storage medium, a processor, and a computer program stored on the computer-readable storage medium and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the power plant equipment voltage heating type defect detection method as described in any one of claims 1-7.

Images