Basin-type insulator superficial defect self-identification method, system and equipment based on infrared thermal wave image and medium

Abstract

The invention discloses a basin-type insulator superficial defect self-recognition method, system and equipment based on an infrared thermal wave image and a medium, and relates to the technical field of nondestructive testing of high-voltage electrical equipment, and the method comprises the following steps: collecting an infrared thermal image sequence containing a thermal response process; preprocessing the acquired infrared thermal image sequence, extracting space, texture and thermal gradient feature information of a superficial defect area in the preprocessed image based on a multi-scale feature fusion strategy, and outputting a defect segmentation result corresponding to the infrared thermal image; inputting a defect segmentation result into a target recognition model with an attention regulation mechanism to complete automatic judgment of defect types and positions; and the defect area and depth estimation value is calculated after discrimination. According to the method, non-contact and intelligent recognition and quantitative analysis of the superficial defect of the basin-type insulator can be realized by constructing a complete process from thermal wave excitation perception, image preprocessing, multi-scale feature extraction, attention mechanism recognition and discrimination to defect parameter quantitative estimation.

Description

Technical Field

[0001] This invention relates to the field of non-destructive testing technology for high-voltage electrical equipment, and in particular to a method, system, device, and medium for self-identification of shallow defects in basin-type insulators based on infrared thermal imaging. Background Technology

[0002] Pot-type insulators, as key components in gas-insulated switchgear (GIS), are widely used in high-voltage power transmission and transformation systems. Their main functions are to provide electrical insulation and mechanical support. However, due to long-term exposure to high-voltage electric fields, mechanical stress, environmental pollution, and aging, shallow defects such as micro-cracks, dirt accumulation, scratches, and bubbles easily develop on the surface of pot-type insulators. If these defects are not detected and addressed in a timely manner, they may lead to a decline in insulation performance and even trigger serious accidents such as flashover and breakdown, threatening the safe operation of the power grid.

[0003] Existing methods for detecting defects in basin-type insulators 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 infrared thermography-based methods for detecting low-value insulators identify defects by extracting grayscale features from the steel cap area. However, this method mainly targets 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 basin-type insulators 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) lack of adaptive self-identification mechanism, and inability to classify defect types in real time.

[0005] Therefore, a novel self-identification technology for shallow defects in basin insulators based on infrared thermal imaging is needed to improve detection accuracy, automation level, and practicality. Summary of the Invention

[0006] In view of the above-mentioned problems, the present invention is proposed.

[0007] Therefore, the problem to be solved by this invention is that traditional image processing algorithms such as edge detection and threshold segmentation are difficult to handle defect morphology in complex backgrounds; they lack an adaptive self-identification mechanism and cannot classify defect types in real time.

[0008] To address the aforementioned technical problems, this invention provides the following technical solution: a method for self-identification of shallow defects in basin-type insulators based on infrared thermal wave images, comprising: applying thermal wave excitation to the basin-type insulator to elicit the response difference between the defect area and the normal area during the thermal diffusion process, and acquiring an infrared thermal image sequence containing the thermal response process; preprocessing the acquired infrared thermal image sequence, extracting spatial, textural, and thermal gradient feature information of the shallow defect area in the preprocessed image based on a multi-scale feature fusion strategy, and outputting a defect segmentation result corresponding to the infrared thermal image; inputting the defect segmentation result into a target recognition model with an attention adjustment mechanism to complete the automatic identification of the defect type and location; and calculating the defect area and depth estimates after identification.

[0009] As a preferred embodiment of the self-identification method for shallow defects in basin-type insulators based on infrared thermal wave images described in this invention, the acquisition of an infrared thermal image sequence including the thermal response process includes: configuring an infrared thermal imager at the position where the excitation device is aligned with the insulator under test, and setting acquisition parameters to cover the complete time sequence from the moment of excitation to the predetermined cooling stage; setting an acquisition time window with the goal of capturing the transient and subsequent cooling process caused by thermal wave excitation, and acquiring frame by frame with a preset time resolution to obtain an infrared thermal image sequence; and saving the acquired infrared thermal image sequence in a specified data format.

[0010] As a preferred embodiment of the self-identification method for shallow defects in basin-type insulators based on infrared thermal wave images described in this invention, the preprocessing of the acquired infrared thermal image sequence includes: performing noise suppression processing on the infrared thermal image sequence to enhance the thermal response difference between the defect area and the background area in the image; optimizing the grayscale distribution of the image based on noise suppression to make the weak defect features visually prominent; performing inter-frame alignment and registration processing to eliminate spatial offset caused by disturbances in the image sequence and ensure the positional consistency of the defect area in multiple frames; and reconstructing the temporal features of the registered infrared thermal image sequence.

[0011] The beneficial effects of this preferred technical solution are as follows: This technical solution effectively improves the clarity and consistency of the image sequence in the subsequent feature extraction process by performing multi-stage preprocessing on the original infrared image sequence, such as image enhancement, denoising, time registration and background suppression.

[0012] As a preferred embodiment of the self-identification method for shallow defects in basin-type insulators based on infrared thermal images described in this invention, the output of defect segmentation results corresponding to the infrared thermal image includes: performing feature encoding on the preprocessed infrared thermal image, constructing a feature expression mechanism with scale awareness, and performing pixel-level feature classification to extract shallow and deep features from the image; aligning the shallow and deep features according to spatial resolution and fusing them; performing edge enhancement processing on the fused multi-scale features, and outputting a defect region segmentation map with the same size as the original image.

[0013] The beneficial effects of this preferred technical solution are: by constructing a multi-scale fusion mechanism, multi-dimensional feature information of the image at the levels of spatial details, texture patterns and thermal gradient distribution can be extracted simultaneously, avoiding the problems of information loss or local texture blurring at a single scale.

[0014] As a preferred embodiment of the self-identification method for shallow defects in basin-type insulators based on infrared thermal wave images described in this invention, the pixel-level feature classification includes: calculating the probability distribution of the defect category and background category to which each pixel belongs; optimizing the segmentation accuracy based on the cross-entropy criterion; and generating a binary mask image through thresholding to form a preliminary defect region segmentation result.

[0015] As a preferred embodiment of the self-identification method for shallow defects in basin-type insulators based on infrared thermal wave images described in this invention, the automatic identification of defect type and location includes: inputting a defect region segmentation map into a target recognition model; employing a feature adjustment mechanism with region perception capability to selectively enhance the spatial features in the input image, guiding the model to focus on the thermal response region; in the model's classification branch, classifying each defect region and outputting a prediction result containing a defect type label and corresponding confidence level; in the model's localization branch, combining thermal response intensity and structural features to regress and locate the spatial position of the defect, outputting a defect bounding box consistent with the original image coordinate system; and performing redundancy processing on the results output from all branches, removing overlapping areas, and generating a final defect detection result list.

[0016] As a preferred embodiment of the self-identification method for shallow defects in basin-type insulators based on infrared thermal images described in this invention, the calculation of defect area and depth estimates includes: extracting the boundary contours corresponding to the defect targets in the image based on the final defect detection result list, mapping them to the infrared thermal image space, and constructing a pixel-level defect region mask; combining the spatial resolution parameters of each pixel in the infrared thermal image, counting the number of pixels in all target regions in the defect mask, and calculating the projected area value of each defect target at its actual size using area conversion rules; extracting the infrared thermal response change curve within the defect region and constructing a temperature response time series within that region; inputting the obtained temperature response time series into a preset thermal-structural response method to estimate the depth range of the defect region and outputting the corresponding depth estimate; and managing and structuring the area and depth estimation results of all defect targets by number, and generating a defect information report.

[0017] To address the aforementioned technical problems, this invention provides the following technical solution: a self-identification system for shallow defects in basin-type insulators based on infrared thermal wave images, comprising: a data acquisition module, a data processing module, a defect identification module, and a parameter calculation module; the data acquisition module is used to apply thermal wave excitation to the basin-type insulator, stimulating the response difference between the defect area and the normal area during the thermal diffusion process, and acquiring an infrared thermal image sequence containing the thermal response process; the data processing module is used to preprocess the acquired infrared thermal image sequence, extract the spatial, texture, and thermal gradient feature information of the shallow defect area in the preprocessed image based on a multi-scale feature fusion strategy, and output the defect segmentation result corresponding to the infrared thermal image; the defect identification module is used to input the defect segmentation result into a target recognition model with an attention adjustment mechanism to complete the automatic discrimination of defect type and location; the parameter calculation module calculates the defect area and depth estimates after discrimination.

[0018] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps of the self-identification method for shallow defects in basin insulators based on infrared thermal wave images as described above.

[0019] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the self-identification method for shallow defects in basin-type insulators based on infrared thermal wave images as described above.

[0020] The beneficial effects of this invention are as follows: By constructing a complete process from thermal wave excitation sensing, image preprocessing, multi-scale feature extraction, attention mechanism identification and discrimination to defect parameter quantification estimation, this invention can achieve non-contact, intelligent identification and quantitative analysis of shallow defects in basin insulators. It not only improves the accuracy and robustness of defect identification, but also supports comprehensive evaluation of defect type, location, area and depth, and has good adaptability, scalability and engineering practical value. Attached Figure Description

[0021] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a flowchart of a method for self-identification of shallow defects in basin-type insulators based on infrared thermal wave images, as described in Example 1.

[0023] Figure 2 This is an image showing the identification results of a self-identification method for shallow defects in basin-type insulators based on infrared thermal wave images in Example 2. Detailed Implementation

[0024] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0025] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0026] Example 1, referring to Figure 1 This is the first embodiment of the present invention, which provides a method for self-identification of shallow defects in basin-type insulators based on infrared thermal wave images, including: S1: Apply thermal wave excitation to the basin insulator to stimulate the response difference between the defect area and the normal area during the thermal diffusion process, and acquire an infrared thermal image sequence containing the thermal response process.

[0027] S2: Preprocess the acquired infrared thermal image sequence, extract the spatial, texture and thermal gradient features of the shallow defect region in the preprocessed image based on the multi-scale feature fusion strategy, and output the defect segmentation result corresponding to the infrared thermal image.

[0028] S3: Input the defect segmentation results into the target recognition model with attention adjustment mechanism to complete the automatic identification of defect type and location.

[0029] S4: Calculate the estimated area and depth of the defect after discrimination.

[0030] It should be noted that currently, in engineering projects, manual inspection or contact instruments (such as ultrasonic and electrical testing methods) are still widely used to detect defects in insulators. This not only has the problems of high labor costs and long response cycles, but is also easily affected by external conditions such as climate and altitude, and poses safety hazards in high-voltage live scenarios. Some recognition methods based on visible light or infrared images rely on single-scale texture or brightness features, which are sensitive to factors such as complex image backgrounds, changes in lighting, and surface contamination, and are prone to misjudgment or missed judgment, making them unsuitable for the complex working environment of actual power grid sites.

[0031] Therefore, to address the aforementioned issues, the following steps (S1-S4) are employed: First, thermal wave excitation is used to induce thermal diffusion differences on the surface of the basin insulator, and infrared image sequences generated during the thermal response process are continuously acquired. Then, preprocessing operations such as enhancement, denoising, and background suppression are performed on the image sequences to highlight potential defect areas. Next, multi-scale fusion is used to extract spatial structure, texture distribution, and thermal gradient features from the images to construct a defect region representation. Furthermore, a recognition model with an attention regulation mechanism is used to automatically determine the location and type of defects. Finally, the area and depth of the defects are quantitatively estimated by combining the thermal response change curve and material properties, outputting structured diagnostic results.

[0032] Example 2, refer to Figure 2 This is the second embodiment of the present invention, which differs from the first embodiment in that: a method for self-identification of shallow defects in basin-type insulators based on infrared thermal wave images further includes, in step S1, the acquisition of an infrared thermal image sequence containing the thermal response process includes the following steps A1-A3: A1: An infrared thermal imager is positioned at the location where the excitation device is aligned with the insulator under test, and the acquisition parameters are set to cover the complete time sequence from the moment of excitation to the predetermined cooling stage.

[0033] A2: Set the acquisition window with the goal of capturing the transient and subsequent cooling process caused by thermal wave excitation, and acquire frame by frame with a preset time resolution to obtain an infrared thermal image sequence.

[0034] A3: Save the acquired infrared thermal image sequence according to the specified data format.

[0035] Specifically, thermal wave excitation is a fundamental step in this invention. Surface defects in basin-type insulators (such as cracks with a depth <1mm) can lead to uneven heat diffusion. Under pulsed thermal wave excitation, the rate of heat accumulation or dissipation in the defect area differs from that in the normal area, creating a thermal contrast.

[0036] Specific implementation: (1) Excitation parameter settings: pulse width is 10-50ms, energy density is 1-5J / cm². For phase-locked mode, excitation frequency f=0.1-1Hz, excitation period T=1 / f.

[0037] (2) Image acquisition: Using an infrared camera, the acquisition temperature range is -20℃ to 150℃, and the acquisition sequence is... ,in For time frames, Represents pixels exist Temperature value at any given time.

[0038] Mathematical Model: By establishing a semi-infinite homogeneous medium heat conduction model, and combining the analysis method of first-order differential highlighting signal curvature change and second-order differential characterizing the characteristics of heat wave propagation acceleration, it was 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 the minimum peak time and the maximum peak time.

[0039] The collected data is pre-stored in MAT format for easy subsequent processing.

[0040] Furthermore, in step S2, the preprocessing of the acquired infrared thermal image sequence includes the following steps B1-B4: B1: Noise suppression processing is applied to the infrared thermal image sequence to enhance the difference in thermal response between the defect area and the background area in the image.

[0041] B2: Based on noise suppression, the grayscale distribution of the image is optimized to make weak defect features stand out visually.

[0042] B3: Perform inter-frame alignment and registration processing to eliminate spatial offsets caused by disturbances in the image sequence and ensure the consistency of the position of defective areas in multiple frames of images.

[0043] B4: Reconstruct the temporal features of the registered infrared thermal image sequence.

[0044] In this embodiment of the application, step S2, the preprocessing employs a multi-stage collaborative enhancement and reconstruction strategy, including the following steps S211-S215: S211: Multi-scale noise suppression processing is performed on the image sequence. A wavelet transform denoising method based on frequency domain analysis is used to separate high-frequency interference components and retain low-frequency structural information.

[0045] S212: Apply local contrast adaptive enhancement methods (such as adaptive histogram equalization) to increase the gray-scale gradient amplitude in the defect hotspot region and enhance the boundary performance of shallow micro-defects.

[0046] S213: An image registration algorithm based on correlation matching is adopted to align all frame references to the reference frame at the initial moment of the thermal response, ensuring that the defect structure has spatial consistency in the sequence.

[0047] S214: Introducing Thermal Signal Reconstruction (TSR) technology, which reconstructs pixel-level sequences based on the thermal diffusion evolution law in the time dimension, thereby enhancing the dynamic thermal anomaly performance of defective areas.

[0048] S215: Calculate the curvature of the thermal response intensity change curve for each pixel, and select the most expressive and clear image frame from the sequence as the subsequent recognition input, using the local curvature peak as the criterion.

[0049] In an optional implementation, the preprocessing may also employ a "lightweight edge enhancement + global frame alignment" method, including the following steps S221-S224: S221: Perform median filtering or guided filtering on the image to quickly suppress noise interference with high discontinuities in the thermal image sequence.

[0050] S222: Employs an edge gradient enhancement filter kernel to enhance the edge response of local areas in the image, highlighting the boundaries of fine-line defects such as cracks and scratches.

[0051] S223: Geometric registration of sequential images is achieved using the weighted feature point method to ensure that the position of the defect area remains consistent in different frames and to avoid recognition errors.

[0052] S224: Select the overall sharpness index of the image frame as the evaluation standard, and select the frame with the best sharpness as the input for subsequent processing.

[0053] In another alternative implementation, the preprocessing may also employ a method of local anomaly detection and multi-reference frame fusion, including the following steps S231-S234: S231: Model the noise distribution of the original heatmap sequence, estimate the local noise mean and standard deviation, and perform adaptive denoising by dynamically adjusting the filter intensity according to the noise level.

[0054] S232: Establish a local thermal anomaly detection mask based on the pixel contrast variation amplitude, automatically identify the preliminary location of defect hot spots, and provide ROI reference for enhancement and extraction areas.

[0055] S233: Image frames are fed into a lightweight frame selection network in a sequential structure, and the variation features of defects in multiple reference frames are fused to generate a fused image to preserve the complete thermal response morphology.

[0056] S234: The fusion result is further fed into the spatial-temporal joint filtering module for final enhancement, outputting clear, high-contrast, and low-background-interference heatmap frames for segmentation model.

[0057] It should be noted that the scheme in this embodiment effectively improves the clarity and consistency of the image sequence in the subsequent feature extraction process by performing multi-stage preprocessing on the original infrared image sequence, such as image enhancement, denoising, time registration and background suppression. Especially in scenarios with complex insulator structures and a lot of background heat source interference, the preprocessing strategy can reduce the impact of artifact noise, highlight the thermal response changes of the defect area, and enable the model to more accurately identify the defect boundary and contour shape, thereby improving the overall stability and robustness of the recognition system.

[0058] Furthermore, in step S2, outputting the defect segmentation result corresponding to the infrared thermal image includes the following steps B5-B7: B5: Perform feature encoding on the preprocessed infrared thermal image, construct a feature representation mechanism with scale awareness, and perform pixel-level feature classification to extract shallow and deep features from the image.

[0059] B6: Align shallow and deep features according to spatial resolution and then fuse them.

[0060] B7: Perform edge enhancement processing on the fused multi-scale features and output a defect region segmentation map with the same size as the original image to achieve pixel-level annotation of shallow defect regions in the image.

[0061] Specifically, performing pixel-level feature classification includes the following steps: Calculate the probability distribution of each pixel to the defect category and the background category.

[0062] Optimize segmentation accuracy based on cross-entropy criterion.

[0063] A binary mask image is generated by thresholding to form a preliminary defect region segmentation result.

[0064] In this embodiment of the application, in step B5, the feature representation mechanism with scale-aware capability adopts scale-aware representation based on a pyramid pooling structure, including the following steps B511-B516: B511: The preprocessed image is passed through a basic convolutional encoder to obtain a high-level feature map F0 (size H×W×C, where H represents height, W represents width, and C represents the number of channels).

[0065] B512: Apply global average pooling operations of different sizes (e.g., 1×1, 2×2, 3×3, 6×6) to F0 to obtain several low-resolution feature maps F1 to F2. 4。

[0066] B513: Perform bilinear interpolation upsampling on the feature maps at each scale to the original size H×W, and align the number of channels by channel-wise convolution (1×1).

[0067] B514: Concatenate the upsampled multi-scale feature maps into a fusion tensor F_concat along the channel dimension.

[0068] B515: Perform convolutional normalization and nonlinear activation on F_concat to form a fused feature F_fuse as a multi-scale output.

[0069] B516: Input F_fuse to the decoder or edge enhancement module for pixel-level segmentation or thermal gradient extraction.

[0070] In an optional implementation, the scale-aware feature representation mechanism can also employ a parallel receptive field expansion structure based on dilated convolutional networks, including the following steps B521-B525: B521: The input feature map F is fed into four parallel convolution branches, and the dilatation rates of each branch are set to d1=1, d2=2, d3=4, and d4=8, respectively.

[0071] B522: The kernel size of each branch is fixed at 3×3, and the receptive field coverage is controlled by the porosity.

[0072] B523: Perform batch normalization and activation function processing on the feature maps output by each branch.

[0073] B524: Concatenate the outputs of the four branches along the channel dimension to form a comprehensive feature tensor F_cat.

[0074] B525: Perform a 1×1 convolution on F_cat to obtain the fused feature F_fuse, which contains both local edge information (low void ratio branch) and global thermal diffusion information (high void ratio branch).

[0075] In another alternative implementation, the scale-aware feature representation mechanism can also employ an attention-guided multi-scale feature fusion structure, including the following steps B531-B536: B531: Establish a multi-scale feature set {F1, F2, F3} between shallow and deep feature outputs, where F1 has high resolution and shallow semantics; F2 has medium resolution and medium semantics; and F3 has low resolution and strong semantics.

[0076] B532: Calculate the channel statistics for each scale feature (such as the result of global average pooling) and generate the channel attention weight vector W_c through a two-layer fully connected structure.

[0077] B533: Channel weighting for each scale feature: , where n∈(1, 2, 3). ⊙ represents the nth channel, and ⊙ represents channel-by-channel multiplication.

[0078] B534: Calculate the spatial saliency mask for each feature map based on a spatial attention mechanism. To highlight areas of thermal anomalies.

[0079] B535: Spatial weighted fusion of multi-scale features: .

[0080] B536: The fused features are fed into the decoding layer or segmentation module to achieve pixel-level defect region segmentation.

[0081] It should be noted that the present invention constructs a multi-scale fusion mechanism to simultaneously extract multi-dimensional feature information of images at the levels of spatial details, texture patterns, and thermal gradient distribution. This avoids the problems of information loss or local texture blurring at a single scale. Multi-scale modeling not only enhances the system's adaptability to shallow defects of different sizes, shapes, and edge blurring, but also improves the consistency of recognition results through feature alignment and fusion, significantly improving the accuracy of defect recognition and the ability to resist background interference.

[0082] Furthermore, in step S3, the automatic identification of defect type and location includes the following steps C1-C4: C1: Input the defect region segmentation map into the target recognition model, and use a feature adjustment mechanism with region perception capability to selectively enhance the spatial features in the input image, guiding the model to focus on the thermal response region.

[0083] C2: In the classification branch of the model, classify each defect region and output a prediction result containing the defect type label and the corresponding confidence level.

[0084] C3: In the localization branch of the model, the spatial location of the defect is regressed by combining the heat map response intensity and structural features, and the defect bounding box is output in a coordinate system consistent with the original image.

[0085] C4: Redundancy processing is performed on the output results of all branches, overlapping areas are removed, and a final defect detection result list is generated. The identification results are as follows: Figure 2 As shown.

[0086] Specifically, the target recognition model used in this embodiment is the improved YOLOv5 (You Only Look Once) model, which integrates CBAM (Convolutional Attention Module) to enhance attention to small defects.

[0087] Model Improvement: Backbone network: CSPDarknet (cross-stage partial connection structure) was replaced with ResNet (residual network) to better handle thermal images.

[0088] Neck: Add BiFPN (Bi-directional Feature Pyramid Network) for multi-scale fusion to improve the detection of shallow defects.

[0089] Head classification: The CIOU loss function is used, with the loss L = L_cls + L_box + L_obj, where L_cls is the classification loss (cross-entropy), L_box is the bounding box regression, and L_obj is the target confidence loss term.

[0090] Training process: The dataset consists of 1000 labeled infrared images of basin insulators (defect types: cracks, dirt, scratches). The Adam (adaptive moment estimation) optimizer is used with a learning rate of 0.001, a batch size of 16, and 200 epochs.

[0091] Inference process: Input a preprocessed image, output bounding box, confidence score and category, with a threshold of 0.5 to filter false positives.

[0092] Evaluation criteria: Accuracy = TP / (TP+FP) = 96%, Recall = TP / (TP+FN) = 94%, mAP0.5 (the overlap threshold between predicted and ground truth boxes is 0.5) = 95.5%, where TP represents the number of samples correctly detected as defects, FP represents the number of samples incorrectly detected as defects, and FN represents the number of samples that were actually defects but were not identified.

[0093] In this embodiment of the application, in step C1, the feature adjustment mechanism with region-aware capability adopts a channel weight adjustment method based on the "compression-excitation mechanism," including the following steps C111-C114: C111: Perform global average pooling on the input feature map in the recognition model to extract the overall response intensity on each feature channel.

[0094] C112: A two-layer fully connected neural network structure is used to perform a non-linear mapping on the pooling results to generate the weight coefficients for each channel.

[0095] C113: Perform a channel-by-channel multiplication operation between the weight vector and the original feature map to form a channel-weighted feature map.

[0096] C114: The final output enhances the new features of "high response channel and background suppression channel" for subsequent defect identification processing.

[0097] In an optional implementation, the feature modulation mechanism with region-aware capabilities can also employ a spatially salient region focusing method based on a "self-attention mechanism," including the following steps C121-C124: C121: Perform pixel-wise average pooling and max pooling operations on the input feature map to obtain the background baseline and abnormal peak of the feature response in the spatial dimension, respectively.

[0098] C122: The two pooling results are concatenated, and then a spatial attention map generation module with a sliding window convolution kernel is used to extract the region saliency mask.

[0099] C123: Perform a pixel-by-pixel multiplication operation between the mask and the original feature map, so that the network retains only the regions in the image with "strong thermal gradient abrupt changes and significant boundaries".

[0100] C124: The output feature map has saliency focusing capability, which effectively improves the expression accuracy of defect targets such as hot spots and dirty edges.

[0101] In another alternative implementation, the feature modulation mechanism with region-aware capabilities can also employ a spatial-channel joint modeling method based on a "global self-attention structure," including the following steps C131-C134: C134: Flatten the feature map of the input image into a two-dimensional matrix and use a set of linear mapping functions to generate three sets of representations: "query vector", "key vector" and "value vector".

[0102] C134: The similarity between feature points is calculated by the dot product of the query vector and the key vector, and the weight matrix is ​​obtained by the normalization function.

[0103] C134: Use the weight matrix to perform a weighted summation of the "value vectors" to obtain an output representation that incorporates information from the entire graph.

[0104] C134: Finally, this output is fused with the original feature map proportionally to achieve joint feature enhancement of the channel dimension and spatial dimension.

[0105] It should be noted that this solution introduces an intelligent recognition model with an attention adjustment mechanism, which can dynamically focus on significant areas related to defects in both spatial and channel dimensions, effectively suppressing redundant background features and increasing the recognition weight of the target area. Compared with traditional methods based on thresholds, fixed templates, or manual features, this mechanism has adaptive learning and transfer capabilities, and can more accurately distinguish different types of defects such as surface cracks, foreign matter attachment, or coating peeling, and output their precise location annotations in the image, thereby significantly improving the intelligence level and generalization ability of automatic recognition.

[0106] Furthermore, in step S4, calculating the estimated defect area and depth includes the following steps: D1-D5: D1: Based on the final defect detection result list, extract the boundary contours corresponding to the defect targets in the image and map them into the infrared thermal image space to construct a pixel-level defect region mask.

[0107] D2: Combining the spatial resolution parameters of each pixel in the infrared thermal image, count the number of pixels in all target areas of the defect mask, and calculate the projected area value of each defect target under the actual size through the area conversion rule.

[0108] D3: Extract the infrared thermal response change curve within the defect area and construct the temperature response time series within that area.

[0109] D4: Input the obtained temperature response time series into the preset thermal-structural response method to estimate the depth range of the defect area and output the corresponding depth estimate.

[0110] D5: Number and structure the area and depth estimation results of all defective targets, and generate defect information reports.

[0111] In this embodiment of the application, step D4, the thermal-structural response method adopts an analytical model method based on thermal diffusion theory, including the following steps D411-D415: D411: Extract the infrared thermal response curve of temperature change over time within the defect area, and record the peak temperature and arrival time of the thermal response for each pixel.

[0112] D412: According to the thermal diffusion equation (in λ is the thermal diffusivity, and λ is the thermal conductivity. (where c is density and c is specific heat capacity) to derive the propagation characteristics of heat waves inside the material.

[0113] D413: Temperature difference between defective and non-defective regions By comparing the evolution over time, the defect depth is estimated by calculating the peak time of the second-order differential curve.

[0114] D414: Combining the thermal conductivity parameters of the material with the duration of the excitation pulse The depth estimate corresponding to the defect is obtained using an empirical formula, which is expressed as follows: in, This represents the estimated depth value corresponding to the defect.

[0115] D415: In the output results, a correspondence table is formed between defect depth and thermal response delay, which is used for defect level assessment.

[0116] In an alternative implementation, the thermal-structural response method can also employ a data-driven thermal response-depth regression method, including the following steps D421-D425: D421: Using multiple standard defect samples with known depths and dimensions as references, establish a multi-dimensional sample database containing "thermal response delay - temperature rise amplitude - defect depth".

[0117] D421: Extract key feature parameters from the infrared thermogram sequence of each sample, such as the rate of temperature rise, settling time, logarithmic rate of temperature change, and local gradient curve morphology.

[0118] D423: Use nonlinear regression methods (such as multilayer fully connected regression networks or support vector regressors) to establish a mapping model between thermal features and defect depth.

[0119] D424: Input the actual detected defect thermal features into this mapping model, and predict the corresponding defect depth estimate.

[0120] D425: Performs residual correction on the model output to reduce the bias caused by differences in sampling resolution.

[0121] In another alternative implementation, the thermal-structural response method can also employ a data-driven thermal response-depth regression method, including the following steps D431-D435: D435: The analytical model method based on thermal diffusion theory is used to perform preliminary depth calculation of the defect region and obtain the theoretical depth value h1.

[0122] D435: Calculate the data-driven depth prediction value h2 using the data-driven thermal response-depth regression method.

[0123] D435: Compare the difference between the two values ​​Δh=|h1-h2|. When the difference exceeds the preset threshold, the hybrid compensation mechanism is automatically triggered.

[0124] D435: The hybrid compensation mechanism calculates the depth estimate corresponding to the defect using a weighted average strategy. h = ω1·h1 + ω2·h2 Wherein, ω1 and ω2 are dynamically adjusted based on the heatmap signal-to-noise ratio and model confidence.

[0125] D435: Output the final integrated depth estimate and record the calculated confidence range.

[0126] It should be noted that this technical solution, by constructing a mapping model from thermal response to physical properties (area, depth), bridges the key link between "image-level recognition" and "parameter-level quantitative evaluation," realizing the quantification and evaluability of defect diagnosis. In particular, in defect depth estimation, it integrates thermal diffusion modeling and data-driven regression mechanisms, which are not only applicable to complex materials and heterogeneous structures, but also effectively address the interference of surface thermal noise fluctuations on the estimation results, providing objective and quantitative basic data support for equipment maintenance strategy formulation and reliability assessment.

[0127] Example 3, the third embodiment of the present invention, differs from the previous two embodiments in that it provides a self-identification system for shallow defects in basin-type insulators based on infrared thermal wave images. The system includes a data acquisition module, a data processing module, a defect identification module, and a parameter calculation module. The data acquisition module applies thermal wave excitation to the basin-type insulator, stimulating the response difference between the defective and normal regions during the thermal diffusion process, and acquires an infrared thermal image sequence containing the thermal response process. The data processing module preprocesses the acquired infrared thermal image sequence, extracts spatial, textural, and thermal gradient features of the shallow defective region in the preprocessed image based on a multi-scale feature fusion strategy, and outputs the defect segmentation result corresponding to the infrared thermal image. The defect identification module inputs the defect segmentation result into a target recognition model with an attention adjustment mechanism to automatically determine the defect type and location. The parameter calculation module calculates the estimated defect area and depth after the determination.

[0128] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, essentially, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0129] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-including system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0130] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0131] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented in combination with any of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0132] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for self-identification of shallow defects in basin-type insulators based on infrared thermal wave images, characterized in that: include, Thermal wave excitation was applied to the basin insulator to elicit the response difference between the defect area and the normal area during the thermal diffusion process, and an infrared thermal image sequence containing the thermal response process was acquired. The acquired infrared thermal image sequence is preprocessed, and the spatial, texture and thermal gradient feature information of shallow defect areas in the preprocessed image is extracted based on a multi-scale feature fusion strategy. The defect segmentation result corresponding to the infrared thermal image is then output. The defect segmentation results are input into a target recognition model with an attention adjustment mechanism to automatically identify the defect type and location. After identification, the estimated values ​​of defect area and depth are calculated.

2. The self-identification method for shallow defects in basin-type insulators based on infrared thermal wave images as described in claim 1, characterized in that: The acquired infrared thermal image sequence, which includes the thermal response process, includes... An infrared thermal imager is positioned at the location where the excitation device is aligned with the insulator under test, and the acquisition parameters are set to cover the complete time sequence from the moment of excitation to the predetermined cooling stage. The acquisition window is set with the goal of capturing the transient and subsequent cooling process caused by thermal wave excitation, and frame-by-frame acquisition is performed with a preset time resolution to obtain an infrared thermal image sequence. Save the acquired infrared thermal image sequence according to the specified data format.

3. The self-identification method for shallow defects in basin-type insulators based on infrared thermal wave images as described in claim 2, characterized in that: The preprocessing of the acquired infrared thermal image sequence includes... Noise suppression processing is applied to the infrared thermal image sequence to enhance the difference in thermal response between the defect area and the background area in the image; Based on noise suppression, the grayscale distribution of the image is optimized to make weak defect features stand out visually. Perform inter-frame alignment and registration processing to eliminate spatial offsets caused by disturbances in the image sequence and ensure the consistency of the position of defective areas in multiple frames of images; Temporal features are reconstructed from the registered infrared thermal image sequence.

4. The self-identification method for shallow defects in basin-type insulators based on infrared thermal wave images as described in claim 3, characterized in that: The defect segmentation results corresponding to the output infrared thermal image include: Feature encoding is performed on the preprocessed infrared thermal image to construct a feature representation mechanism with scale awareness, and pixel-level feature classification is performed to extract shallow and deep features from the image. The shallow and deep features are aligned according to spatial resolution and then fused. Edge enhancement processing is performed on the fused multi-scale features to output a defect region segmentation map with the same size as the original image.

5. The self-identification method for shallow defects in basin-type insulators based on infrared thermal wave images as described in claim 4, characterized in that: The pixel-level feature classification includes, Calculate the probability distribution of each pixel to the defect category and the background category; Optimize segmentation accuracy based on cross-entropy criterion; A binary mask image is generated by thresholding to form a preliminary defect region segmentation result.

6. The self-identification method for shallow defects in basin-type insulators based on infrared thermal wave images as described in claim 5, characterized in that: The automatic identification of defect type and location includes, The defect region segmentation map is input into the target recognition model, and a feature adjustment mechanism with region perception capability is used to selectively enhance the spatial features in the input image, guiding the model to focus on the thermal response region. In the classification branch of the model, the category of each defect region is determined, and the prediction result containing the defect type label and the corresponding confidence level is output. In the localization branch of the model, the spatial location of the defect is regressed by combining the heat map response intensity and structural features, and the defect bounding box is output in a coordinate system consistent with the original image. Redundancy is processed for all branch outputs, overlapping areas are removed, and a final list of defect detection results is generated.

7. The self-identification method for shallow defects in basin-type insulators based on infrared thermal wave images as described in claim 6, characterized in that: The calculated estimated defect area and depth include, Based on the final defect detection result list, the boundary contours corresponding to the defect targets in the image are extracted and mapped to the infrared thermal image space to construct a pixel-level defect region mask. By combining the spatial resolution parameters of each pixel in the infrared thermal image, the number of pixels in all target areas of the defect mask is counted, and the projected area value of each defect target under the actual size is calculated by the area conversion rule. Extract the infrared thermal response change curves within the defect area and construct the temperature response time series within that area; The obtained temperature response time series is input into the preset thermal-structural response method to estimate the depth range of the defect region and output the corresponding depth estimate. The area and depth estimation results of all defective targets are numbered, managed, and structured, and defect information reports are generated.

8. A self-identification system for shallow defects in basin-type insulators based on infrared thermal wave images, employing the self-identification method for shallow defects in basin-type insulators based on infrared thermal wave images as described in any one of claims 1 to 7, characterized in that: It includes a data acquisition module, a data processing module, a defect identification module, and a parameter calculation module; The data acquisition module is used to apply thermal wave excitation to the basin insulator, stimulate the response difference between the defect area and the normal area during the thermal diffusion process, and acquire an infrared thermal image sequence containing the thermal response process. The data processing module is used to preprocess the acquired infrared thermal image sequence, extract the spatial, texture and thermal gradient feature information of the shallow defect area in the preprocessed image based on the multi-scale feature fusion strategy, and output the defect segmentation result corresponding to the infrared thermal image. The defect identification module is used to input the defect segmentation results into a target identification model with an attention adjustment mechanism to automatically identify the defect type and location. The parameter calculation module calculates the estimated area and depth of the defect after discrimination.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the self-identification method for shallow defects in basin-type insulators based on infrared thermal wave images, as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the self-identification method for shallow defects in basin-type insulators based on infrared thermal wave images, as described in any one of claims 1 to 7.

Images