Insulation pull rod superficial defect detection method and system based on infrared thermal wave image

Abstract

This invention discloses a method and system for detecting shallow defects in insulating tie rods based on infrared thermal wave images, belonging to the field of non-destructive testing technology for high-voltage electrical equipment. The method includes: applying thermal wave excitation to the surface of the insulating tie rod and acquiring a sequence of infrared thermal images; preprocessing the acquired infrared thermal image sequence; segmenting the defect region using a multi-scale feature fusion network; classifying and locating the defects using a deep learning model with an integrated attention mechanism, and optionally performing quantitative evaluation of the defect area and depth. The system includes a thermal wave excitation module, an image acquisition module, an image preprocessing module, a feature extraction module, and a recognition and classification module. This invention combines infrared thermal wave technology with deep learning algorithms, effectively improving the detection accuracy and automation level of shallow defects. It has advantages such as high recognition accuracy, comprehensive functions, and strong practicality, and is suitable for condition monitoring and intelligent operation and maintenance of power equipment.

Description

Technical Field

[0001] This invention relates to the field of non-destructive testing technology for high-voltage electrical equipment, specifically to a method and system for detecting shallow defects in insulating tie rods based on infrared thermal imaging. Background Technology

[0002] Insulating tie rods, as key components in gas-insulated switchgear (GIS), play a crucial role in mechanical operation and electrical isolation. The integrity of their insulation performance directly impacts the safe and stable operation of the power grid. However, during manufacturing or operation, micron- to millimeter-sized air gap defects may form near the surface of the insulating tie rod. These defects, under strong electric fields, can lead to electric field distortion and partial discharge, causing continuous degradation of the insulation material and potentially resulting in catastrophic insulation failures. Therefore, early and accurate detection of shallow defects in insulating tie rods is a critical step in preventing accidents and enabling condition-based maintenance.

[0003] Existing methods for detecting defects in insulating tie rods mainly include visual inspection, ultrasonic testing, X-ray inspection, and infrared thermography. Visual inspection is highly subjective and inefficient; ultrasonic and X-ray testing require contact with equipment, are complex to operate, and pose radiation risks. Infrared thermography, as a non-contact, non-destructive testing technology, has been preliminarily applied to insulator defect detection. For example, existing low-value insulator detection methods based on infrared thermography identify defects by extracting grayscale features from the steel cap area. However, this method is mainly applicable to porcelain insulator strings and lacks sufficient sensitivity to shallow defects, making automatic identification impossible. Furthermore, some studies have used pulsed infrared thermal wave technology to detect internal defects in composite insulators, but algorithms for identifying shallow defects in insulating tie rods still have limitations, such as inaccurate feature extraction, poor real-time performance, and sensitivity to noise interference.

[0004] The shortcomings of the existing technology are mainly reflected in: (1) the thermal response signal of shallow defects is weak and the image signal-to-noise ratio is low, resulting in low detection accuracy; (2) traditional image processing algorithms such as edge detection and threshold segmentation are difficult to handle defect morphology in complex backgrounds; (3) there is a lack of adaptive self-identification mechanism, and it is impossible to classify defect types in real time; (4) in the authorized patent, such as the low-value insulator detection method based on infrared thermal imaging described in CN104267063A, it is limited to gray-scale region extraction and does not involve the integration of deep learning algorithms, so it cannot adapt to the curved surface structure and multiple types of defects of the insulating rod.

[0005] Therefore, there is an urgent need for a new detection method for shallow defects in insulating tie rods to improve detection accuracy, automation level and practicality. Summary of the Invention

[0006] The technical problem to be solved by this invention is to address the shortcomings of the prior art by proposing a method for detecting shallow defects in insulating tie rods based on infrared thermal imaging, specifically including the following steps: S1. Thermal Excitation and Image Acquisition: Apply thermal excitation to the surface of the insulating tie rod and use an infrared camera to acquire a sequence of infrared thermal images during the thermal excitation process; S2. Image preprocessing: Perform preprocessing operations on the infrared thermal image sequence acquired in S1; S3. Defect Feature Extraction: The preprocessed infrared thermal image sequence is input into a multi-scale feature fusion network to segment shallow defect regions; S4. Defect Identification and Classification: The segmented defect region features are input into a deep learning recognition model with an integrated attention mechanism to classify the defect region features and output the type and location information of the defect.

[0007] Furthermore, in step S1, the heat excitation adopts a pulsed / phase-locked loop heat wave source.

[0008] Furthermore, in step S2, image preprocessing includes wavelet transform denoising and adaptive histogram equalization enhancement.

[0009] Furthermore, in step S3, the multi-scale feature fusion network is an infrared thermal image semantic segmentation network built based on ResNet50 and DeepLabV3+.

[0010] Furthermore, the infrared thermal image semantic segmentation network sets up parallel convolutional branches with different hole rates to capture multidimensional features, restores the spatial resolution of the image through bilinear upsampling and convolution operations, and outputs a semantic segmentation map of the same size as the original infrared thermal image sequence.

[0011] Furthermore, in step S4, the deep learning recognition model integrating the attention mechanism is a YOLOv5 learning model integrating the CBAM attention mechanism.

[0012] Furthermore, following S4 defect identification and classification, step S5 defect quantification assessment is also included: calculating the area of ​​the defect using the thermal radiation intensity curve acquired by the infrared camera, and estimating the depth of the defect based on the thermal radiation intensity and acquisition time.

[0013] Furthermore, a shallow defect detection system for insulating tie rods based on infrared thermal imaging is proposed to implement a method for detecting shallow defects in insulating tie rods based on infrared thermal imaging, including: Infrared thermal wave excitation module applies thermal wave excitation to the surface of the insulating tie rod; The image acquisition module acquires a sequence of infrared thermal images during the thermal excitation process; The image preprocessing module performs denoising, enhancement, and registration processing on the infrared thermal image sequence; The defect feature extraction module has a built-in multi-scale feature fusion network to extract defect features and perform shallow defect region segmentation. The self-identification and classification module has a built-in deep learning recognition model with an integrated attention mechanism. It classifies and identifies the segmented defect features and outputs the defect type and location.

[0014] Furthermore, the image acquisition module includes an infrared camera with a resolution of no less than 640×512 pixels and a thermal sensitivity better than 50mK.

[0015] Furthermore, the system also includes a defect quantification assessment module, which quantifies the area and depth of defects based on the segmentation results.

[0016] Compared with the prior art, the present invention achieves the following technical effects: 1. High detection accuracy: Taking into account the weak thermal signals of shallow defects, the system achieves an accuracy rate of over 95% through multi-scale feature fusion and attention mechanism, significantly improving detection sensitivity.

[0017] 2. High degree of automation: It realizes full-process automation from image acquisition to defect identification, classification and quantitative evaluation, which greatly reduces manual intervention and improves detection efficiency and consistency.

[0018] 3. Comprehensive functions: It not only completes the judgment and location of defects, but also realizes the automatic classification of defect types and the quantitative assessment of size (area, depth), providing a more reliable decision-making basis for equipment condition maintenance.

[0019] 4. High practicality: The algorithm design balances accuracy and efficiency, has a fast processing speed, and can be integrated into the online monitoring system of GIS equipment or used in inspection equipment, with broad engineering application prospects. Attached Figure Description

[0020] For ease of explanation, the present invention will be described in detail below with reference to specific embodiments and accompanying drawings.

[0021] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a block diagram of the overall system structure of the present invention; Figure 3 The original infrared thermogram sequence of the surface of the dry-stretched sample containing defective insulation; Figure 4 This is a schematic diagram of an image preprocessing algorithm; Figure 5 Network structure diagram for defect feature extraction; Figure 6 This is an example image showing the experimental verification results. Detailed Implementation

[0022] The following are specific embodiments of the present invention, described in conjunction with the accompanying drawings, to further illustrate the technical solutions of the present invention. However, the present invention is not limited to these embodiments. Specific details, such as particular configurations, are provided in the following description merely to aid in a comprehensive understanding of the embodiments of the present invention. Therefore, those skilled in the art should understand that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of the present invention.

[0023] It should be noted that, unless otherwise specified, the embodiments and features described in this invention can be combined with each other.

[0024] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Figure 1 As shown, the method for detecting shallow defects in insulating tie rods based on infrared thermal imaging of the present invention includes the following steps: S1. Thermal Excitation and Image Acquisition: A pulsed or phase-locked loop thermal wave source is used to apply thermal excitation to the surface of the insulating tie rod, stimulating the thermal diffusion difference in the defect area. A high-resolution infrared camera is used to acquire a sequence of thermal images. The heat source for applying thermal excitation is a halogen lamp with a power of 5000W and a pulse width of 7s. The acquisition time is set to 30s, and the acquired thermal image sequence is the thermal imaging sequence within 30s of the cooling phase after the completion of thermal excitation.

[0025] S2. Image Preprocessing: The infrared thermal image sequence acquired in S1 is subjected to denoising, contrast enhancement, and multi-frame registration to eliminate environmental noise and motion blur. Specifically, during the generation and processing of the thermal images, enhancement is performed only on predefined key areas to avoid noise amplification caused by full-image enhancement, highlighting the contrast of defect areas and making the identification of subsequent frames with optimal clarity more accurate. This method is suitable for identifying minute defects or low-contrast features in infrared thermal images.

[0026] S3. Defect Feature Extraction: The preprocessed infrared thermal image sequence is input into a multi-scale feature fusion network to extract shallow features (texture and edges of shallow defects) and deep features (semantic features), which are then fused to segment the shallow defect region. Deep features are enhanced by dilated convolutions with different dilation rates to increase the receptive field of the convolutional network. The sensed depth features are stacked in a pyramid structure and convolutionally processed to obtain a superimposed feature layer. The multi-scale feature fusion network is an infrared thermal image semantic segmentation network built on ResNet50 and DeepLabV3+. It achieves multi-dimensional feature capture by setting parallel convolutional branches with different dilation rates, and improves edge localization accuracy by restoring the image spatial resolution through bilinear upsampling and convolution operations. Finally, it outputs a semantic segmentation map of the same size as the original image, achieving pixel-level defect region annotation. DeepLabV3+, as an improved version of DeepLabV3, introduces a dilated spatial pyramid module to capture multi-scale contextual information.

[0027] S4. Defect Identification and Classification: The segmented defect region features are input into the YOLOv5 learning model with integrated CBAM attention mechanism to classify the defect region features and output the type and location information of the defects.

[0028] The overall system block diagram provided by this invention is as follows: Figure 2 As shown, it mainly includes an infrared thermal wave excitation module, an image acquisition module, an image preprocessing module, a defect feature extraction and segmentation module, a self-identification and classification module, and an optional defect quantification and evaluation module.

[0029] Specifically, the system of the present invention includes a hardware component and a software component. The hardware component includes: Thermal Excitation Module: Used to apply thermal excitation to the surface of the insulated tie rod, stimulating the difference in thermal diffusion between the defective and normal areas. A pulsed or phase-locked loop thermal source, such as a high-power halogen lamp, is preferred. Example technical parameters: Power 5000W, adjustable pulse width (e.g., 7s), energy output range 100-500J, phase-locked loop excitation frequency range 0.1-1Hz.

[0030] Image acquisition module: Used to acquire infrared thermal image sequences of the insulated tie rod during thermal excitation and cooling processes. It employs a high-resolution, high-thermal-sensitivity uncooled infrared focal plane camera. Example technical parameters: resolution 640×512 pixels, thermal sensitivity less than 50mK, frame rate 50Hz, spectral response range 8-14μm, temperature measurement range -20℃ to 150℃.

[0031] Control and processing computer: Used to control the synchronization of excitation and acquisition, and to run subsequent image processing and defect recognition software. It is equipped with an image acquisition card and corresponding data interface.

[0032] The software component includes: an image preprocessing module, which performs denoising, enhancement, and registration on the acquired raw infrared thermal image sequence to improve image quality and signal-to-noise ratio; a defect feature extraction module, which incorporates a multi-scale feature fusion network to extract defect features and segment shallow defect regions; and a self-identification and classification module, which automatically classifies and locates the extracted features, identifies the defect type, and outputs its location information. The hardware technical parameters are shown in Table 1.

[0033] Table 1 Technical Specifications

[0034] Preferably, thermal wave excitation is a fundamental step of this invention. Surface defects in the insulating tie rod can lead to uneven heat diffusion. Under pulsed thermal wave excitation, the rate of heat accumulation or dissipation in the defective area differs from that in the normal area, creating a thermal contrast.

[0035] The thermal wave excitation is implemented as follows: 1. Excitation method selection: In pulse mode, set the pulse width to 10-50ms and the energy density to 1-5J / cm². For phase-locked loop (PLL) mode, the excitation frequency f=0.1. -1 Hz, excitation period T=1 / f.

[0036] 2. Image Acquisition: A FLIR A615 infrared camera was used to acquire images within a temperature range of -20℃ to 150℃. After thermal excitation was applied, a series of thermal images of the cooling process on the surface of the insulating tie rod were continuously acquired. The acquisition time series was set as I(t), where t represents the time frame (i.e., the temporal order of acquisition). Then, I(x, y, t) represents the surface temperature value of the insulating tie rod detected at pixel position (x, y) in the image at time frame t. The acquired data was pre-stored in MAT format for subsequent processing. Figure 3 An example of a sequence of raw infrared thermal images of the surface of a defective insulating tie rod is shown.

[0037] By approximating the insulated tie rod as a "semi-infinite homogeneous medium," a heat conduction model of this medium is established to suit the structural characteristics and thermal diffusion scenario of the insulated tie rod. Combining the analysis methods of first-order differential highlighting signal curvature changes and second-order differential characterizing the acceleration characteristics of heat wave propagation, it is found that the logarithmic time second-order differential signal of logarithmic thermal radiation intensity not only responds synchronously with the defect thermal resistance effect, but its peak time also maintains a quantitative correlation with the air gap depth. The peak time of the second-order differential curve includes both minimum and maximum peak times. The heat conduction model equation is as follows: T / t=α ²T+Q(x,y,t) Where α is the thermal diffusivity and Q is the heat source. At the defect, α decreases, leading to an increase in the temperature gradient ΔT. The heat conduction equation is the standard form of Fourier's law of heat conduction, suitable for describing the thermal diffusion process of an insulated tie rod as a "semi-infinite homogeneous medium."

[0038] Preferably, infrared images are susceptible to noise, ambient temperature, and camera shake, thus requiring preprocessing to improve the signal-to-noise ratio. Addressing the issues of low signal-to-noise ratio and blurred image details in the original thermal image sequences, especially when defect sizes are small or hot spot contrast is indistinct, noise interference can easily mask the true signal, severely impacting the reliability of image analysis and subsequent automatic segmentation. Therefore, thermal signal reconstruction (TSR), denoising, filtering, and image enhancement were performed on the thermal image sequences. After processing, the hot spots in the defect area are clearer and their boundaries are more prominent, providing a high-quality thermal image foundation for subsequent research.

[0039] Preferably, the core aspect of this invention lies in feature extraction. This invention employs a multi-scale fusion network based on residual attention to extract shallow features (texture and edges of shallow defects) and deep features (semantic features), and then fuses the two to segment the shallow defect region. A schematic diagram of the infrared image defect segmentation process is shown below. Figure 4 As shown.

[0040] The multi-scale feature fusion network is an infrared thermal image semantic segmentation network built on ResNet50 and DeepLabV3+. This is a defect segmentation scheme that combines deep residual learning with a hollow spatial pyramid structure. It achieves multi-dimensional feature capture by setting parallel convolutional branches with different hole rates, and improves edge localization accuracy by restoring image spatial resolution through bilinear upsampling and convolution operations. Finally, it outputs a semantic segmentation map of the same size as the original image, achieving pixel-level defect region annotation. The defect extraction network structure is as follows: Figure 5 As shown, it specifically includes: The backbone network uses ResNet50 as the core to extract multi-layer feature maps (F1 to F4) from the input image. Shallow feature maps (such as F1) have high resolution and are rich in details such as edges and textures; deep feature maps (such as F4) have low resolution but are rich in semantic information. ResNet50 is used as the backbone network for deep feature extraction, and its residual structure is used to solve the degradation problem of deep networks, extracting feature maps with strong semantic information.

[0041] The fusion module incorporates the dilated spatial pyramid pooling module from the DeepLabV3+ architecture. By parallelizing convolutions with different dilation rates, it expands the receptive field without sacrificing resolution, capturing multi-scale contextual information and effectively identifying shallow defects of varying sizes. Furthermore, through feature fusion and upsampling, the deep contextual features extracted by the dilated spatial pyramid module are upsampled to the same resolution as the shallow feature maps using bilinear interpolation. Subsequently, the upsampled deep features are concatenated with the corresponding shallow features and then fused using a 1×1 convolution to obtain a fused feature F that contains both details and semantics. fused The fusion formula is: F fused =Conv(Concat(Up(F deep ),F shallow )) Where Conv is a 1x1 convolution, F deep This is a deep feature map, output from the bottom layer of ResNet50, containing rich semantic information. shallow This is a shallow feature map, output by the higher layers of ResNet50, containing rich texture and edge details. `Up` is an upsampling operation used to adjust the resolution of the deeper feature map to match that of the shallower one. Figure 1 To.

[0042] The attention mechanism integrates a convolutional block attention module (CBAM) during feature fusion. This module sequentially computes the channel attention map M. c， Used for recalibrating channel weights, and spatial attention map M s， This is used to recalibrate the spatial location weights, thereby enhancing the network's attention to the weak hot spot features on the surface of the insulated tie rod, suppressing background noise interference, and improving segmentation accuracy. These weights are then applied to the fusion features to obtain the final feature F. att F att This is the final feature map after being weighted by the attention mechanism.

[0043] The channel attention formula is as follows: M c =σ(MLP(AvgPool(F)+MaxPool(F))) The formula for spatial attention is: M s =σ{Conv7x7[Concat(AvgPool(F),MaxPool(F))]} The final feature is represented as: F att =F fused ×M c ×M s Among them, Ffused The fusion feature is the fusion feature output by the multi-scale feature fusion network (an infrared thermal image semantic segmentation network built on ResNet50 and DeepLabV3+), where σ is the activation function used to map the weights to the 0-1 interval.

[0044] Channel attention mechanism is used to identify which channel features are more important, thereby enhancing the response of key features; spatial attention mechanism focuses on judging which spatial features are more discriminative. The synergy of the two can enable the network to focus more accurately on defective regions.

[0045] Through a series of convolution and upsampling operations, a semantic segmentation map of the same size as the original image is finally output to achieve pixel-level defect region annotation. Each pixel is labeled as either "background" or "defect," enabling precise pixel-level localization of defect regions. The segmented defect region features are then input into a YOLOv5 learning model that integrates the CBAM attention mechanism.

[0046] The key features of thermal gradient G, local contrast C, and area A after morphological opening operation in the defect features are calculated. The "thermal physical differences" and "spatial morphological features" of shallow defects are extracted from infrared thermal images in a multi-dimensional and accurate manner, providing a highly discriminative and interference-resistant effective input for defect identification, classification and quantitative evaluation of subsequent deep learning models.

[0047] Where, thermal gradient G= , I / x、 I / y represents the partial derivatives of temperature (grayscale) in the horizontal and vertical directions of the infrared thermal image, respectively, and represents the rate of temperature change. The thermal gradient is used to quantify the "spatial temperature change rate difference" between the defect edge and the normal area. By capturing the temperature change characteristics in the x and y directions, the blurred defect edge contours are highlighted from the thermal image, providing clear boundary criteria for pixel-level segmentation and avoiding recognition ambiguity caused by the fusion of defects and background.

[0048] Local contrast C=(I def -I bg ) / I b g, I def The average thermal radiation intensity of the defect region, I b g represents the average thermal radiation intensity of the surrounding normal area. The local contrast is calculated by normalizing the relative gray level difference to offset the intensity interference caused by ambient temperature fluctuations and camera noise, thereby enhancing the thermal radiation difference between the defect area and the surrounding background and making the hot spot characteristics of the tiny defects more prominent.

[0049] The area A of the region after morphological opening is used to remove tiny noise points in the thermal image and smooth the boundary of the defect region, so as to avoid noise being misjudged as the defect. The area of ​​the region after opening can reflect the actual physical size of the defect (rather than a false area containing noise), and directly provide basic data for the quantitative assessment of defects.

[0050] Preferably, the defect area is manually delineated and classified using LabelMe software. Pixels in the image are divided into two categories: background and defects. After the defect is circled, the label indicates the type as "defect". Unlabeled areas are considered background by default. The result after standardization is a JSON file. The JSON file is then converted into a PNG file that can be recognized by the neural network and compared with the original infrared thermal image. Figure 1 The same dataset is used as the training dataset and fed into the deep learning recognition model for training. The deep learning recognition model uses an improved YOLOv5 model, integrating CBAM to enhance its focus on small defects. Improvements include: Backbone network: The original YOLOv5 CSPDarknet backbone network was replaced with ResNet to better handle thermal images.

[0051] Neck network: A bidirectional feature pyramid network (BiFPN) is used to enhance multi-scale feature fusion capabilities and improve shallow defect detection.

[0052] Loss function: The CIoU loss function is used to optimize bounding box regression and improve localization accuracy. The loss is L=L cls +L box +L obj L cls For classification loss (cross entropy), L box For bounding box regression loss, L obj The target confidence loss.

[0053] Among them, the classification loss L cls The Cross-Entropy Loss function is used to measure the loss between the model's predicted defect category probability distribution and the true category label. The formula is expressed as:

[0054] N represents the total number of defect categories (such as cracks, dirt, scratches, etc.). One-hot encoding of the true category (1 indicates that the category is a true label, and 0 indicates that it is not the category). This represents the probability of the predicted category by the model. This formula is used to optimize the model's ability to classify defect types, ensuring that the predicted category is consistent with the true category.

[0055] Bounding box regression loss L box To adopt C IoUThe (Complete Intersection over Union) loss function is used to optimize the matching degree of position (center point coordinates) and size (width, height) between the predicted bounding box and the ground truth bounding box, thereby improving the accuracy of defect localization. The formula is expressed as:

[0056] IoU is the Intersection over Union (IoU) ratio between the predicted bounding box and the ground truth bounding box, measuring the degree of overlap; ρ is the Euclidean distance between the centers of the predicted and ground truth bounding boxes; C is the diagonal length of the smallest bounding rectangle containing both the predicted and ground truth bounding boxes; α is a weighting coefficient used to balance the weights of the center distance and IoU terms; v is the aspect ratio penalty term, defined as... ,in , represents the width and height of the ground truth bounding box, and w and h represent the width and height of the predicted bounding box, used to penalize the aspect ratio difference between the predicted and ground truth bounding boxes. CIoU loss not only considers the degree of overlap (IoU) but also introduces center point distance and aspect ratio constraints, comprehensively optimizing bounding box regression, and is especially suitable for the accurate localization of shallow and small defects in insulating tie rods.

[0057] Target confidence loss L obj The binary cross-entropy loss function is used to determine whether the predicted bounding box contains the target (i.e., whether it is a positive sample). The formula is expressed as: in, The true confidence level is 1 (1 indicates that the predicted bounding box contains the target, and 0 indicates that it does not). This represents the confidence level (a probability value between 0 and 1) of the model's predictions. This optimizes the model's ability to determine "whether a defect is present," filtering out invalid prediction boxes and improving detection efficiency.

[0058] By combining the aforementioned loss functions, the model can simultaneously optimize classification, localization, and confidence prediction, achieving high-precision detection of shallow defects in insulating tie rods. The introduction of the CIoU loss function significantly improves the accuracy of bounding box regression, addressing the insufficient sensitivity of traditional IoU loss to center point distance and aspect ratio, making it particularly suitable for locating minute defects.

[0059] Training and Output: The model was trained using a prepared dataset consisting of 1000 labeled infrared images of insulating tie rods (defect types: cracks, dirt, scratches). The Adam optimizer was used with an initial learning rate of 0.001, a batch size of 16, and 200 training epochs. After training, the preprocessed images were input into the trained improved YOLOv5 model. The model directly outputs the defect category label (e.g., crack, dirt, scratch), confidence score, and its bounding box location in the image (center coordinates, width, height). Low-probability predictions were filtered out by setting a confidence threshold (e.g., 0.5). Figure 5 This section presents a comparison example between the original image and the model's recognition results.

[0060] Optionally, after detection using this invention, a quantitative assessment of defects can also be performed. For the identified defects, further quantitative analysis is conducted to provide a basis for condition assessment.

[0061] Defect area estimation: In the thermal radiation intensity curve, after finding the maximum intensity peak corresponding to the defect hot spot, take 50% of the peak value as the baseline and determine the two points where the baseline intersects the curve. The distance between the two points is the half-width value corresponding to the hot spot area. The thermal radiation intensity curve is a curve collected by an infrared camera with "thermal radiation intensity" as the vertical axis and "spatial position / time" as the horizontal axis. The defect diameter is calculated by the half-width value (HWHM) of the hot spot, that is, by formula (1): (1) Where D is the defect diameter, S HWHM This represents the half-width value of the hot spot. During the defect quantification assessment stage, to quickly assess the severity of the defect, the hot spot area detected on the infrared thermogram is approximated as a circle, and the defect area is estimated using the formula for the area of ​​a circle. This approximate calculation can meet the needs of rapid classification and assessment of shallow defects in insulating tie rods in engineering applications.

[0062] Defect Depth Estimation: Based on the theory of heat wave conduction, there is a quantitative relationship between defect depth and the phase or peak time of the thermal response. Specifically, this is achieved by using the second-order differential curve of logarithmic thermal radiation intensity versus logarithmic time, and extracting the minimum peak time t from this curve. PSDT The defect depth h can be estimated using formula (2): (2) Where h is the defect depth, t PSDT The minimum peak time is the second-order differential curve of logarithmic thermal radiation intensity versus logarithmic time. The thermal diffusivity of the insulating tie rod material is expressed in square meters per second (m²). 2 / s).

[0063] For example, if an insulating tie rod sample is found to have a crack area of ​​2.5 cm² and a depth of 0.8 mm, it is judged to be a moderate defect.

[0064] Experimental verification and results In a laboratory environment, GIS insulated tie rods were simulated, and artificial defects (crack width 0.5 mm, contamination thickness 0.2 mm) were introduced. 100 sets of data were collected and tested using the detection method and system of this invention, yielding the following experimental results.

[0065] Accuracy: Precision, recall, and mean precision (mAP@0.5) were used as evaluation metrics. Experimental results show that the proposed method achieves a precision of over 96%, a recall of over 94%, and an mAP of 95.5% for detecting shallow defects.

[0066] Efficiency: The processing time for a single frame image (including preprocessing, feature extraction, and recognition) is less than 1 second, meeting the needs of real-time or rapid inspection, and improving accuracy by 15%.

[0067] Robustness: Compared with traditional edge detection or threshold segmentation methods, this method exhibits stronger robustness and higher detection accuracy under noise interference.

[0068] In summary, the system and method provided by this invention, through optimized thermal wave excitation and image acquisition schemes combined with advanced deep learning image processing and recognition algorithms, achieve high-precision, high-efficiency, and automated detection of shallow defects in insulating tie rods, providing an effective technical means for intelligent operation and maintenance of power equipment.

[0069] Those skilled in the art to which this application pertains may make various modifications or additions to the specific embodiments described, or adopt similar methods to replace them, without departing from the inventive concept of this application or exceeding the scope defined by the appended claims.

Claims

1. A method for detecting shallow defects in insulating tie rods based on infrared thermal imaging, characterized in that, Includes the following steps: S1. Thermal Excitation and Image Acquisition: Apply thermal excitation to the surface of the insulating tie rod and use an infrared camera to acquire a sequence of infrared thermal images during the thermal excitation process; S2. Image preprocessing: Perform preprocessing operations on the infrared thermal image sequence acquired in S1; S3. Defect Feature Extraction: The preprocessed infrared thermal image sequence is input into a multi-scale feature fusion network to segment shallow defect regions; S4. Defect Identification and Classification: The segmented defect region features are input into a deep learning recognition model with an integrated attention mechanism to classify the defect region features and output the type and location information of the defect.

2. The method for detecting shallow defects in insulating tie rods based on infrared thermal imaging according to claim 1, characterized in that, In step S1, the heat broadcasting excitation adopts a pulsed / phase-locked loop heat wave source.

3. The method for detecting shallow defects in insulating tie rods based on infrared thermal imaging according to claim 1, characterized in that, In step S2, the image preprocessing includes wavelet transform denoising and adaptive histogram equalization enhancement.

4. The method for detecting shallow defects in insulating tie rods based on infrared thermal imaging according to claim 1, characterized in that, In step S3, the multi-scale feature fusion network is an infrared thermal image semantic segmentation network built based on ResNet50 and DeepLabV3+.

5. The method for detecting shallow defects in insulating tie rods based on infrared thermal imaging according to claim 4, characterized in that, The infrared thermal image semantic segmentation network sets up parallel convolutional branches with different hole rates to capture multidimensional features, restores the spatial resolution of the image through bilinear upsampling and convolution operations, and outputs a semantic segmentation map of the same size as the original infrared thermal image sequence.

6. The method for detecting shallow defects in insulating tie rods based on infrared thermal imaging according to claim 1, characterized in that, In step S4, the deep learning recognition model with integrated attention mechanism is a YOLOv5 learning model with integrated CBAM attention mechanism.

7. The method for detecting shallow defects in insulating tie rods based on infrared thermal imaging according to claim 1, characterized in that, Following S4 defect identification and classification, step S5 defect quantification assessment is also included: calculating the area of ​​the defect using the thermal radiation intensity curve acquired by the infrared camera, and estimating the depth of the defect based on the thermal radiation intensity and acquisition time.

8. A shallow defect detection system for insulating tie rods based on infrared thermal imaging, used to implement the method according to any one of claims 1-7, characterized in that, include: Infrared thermal wave excitation module applies thermal wave excitation to the surface of the insulating tie rod; The image acquisition module acquires a sequence of infrared thermal images during the thermal excitation process; The image preprocessing module performs denoising, enhancement, and registration processing on the infrared thermal image sequence; The defect feature extraction module has a built-in multi-scale feature fusion network to extract defect features and perform shallow defect region segmentation. The self-identification and classification module has a built-in deep learning recognition model with an integrated attention mechanism. It classifies and identifies the segmented defect features and outputs the defect type and location.

9. A shallow defect detection system for insulating tie rods based on infrared thermal imaging according to claim 8, characterized in that, The image acquisition module includes an infrared camera with a resolution of not less than 640×512 pixels and a thermal sensitivity better than 50mK.

10. A shallow defect detection system for insulating tie rods based on infrared thermal imaging according to claim 8, characterized in that, The system also includes a defect quantification assessment module, which quantifies the area and depth of defects based on the segmentation results.

Images