Nondestructive testing method and system for power transmission line

Abstract

The invention relates to the technical field of power transmission line detection, in particular to a nondestructive detection method and system for a power transmission line. The unmanned aerial vehicle is used for putting the airborne nondestructive testing robot to the power transmission line through the special mounting device, and the unmanned aerial vehicle is adaptive to various line structures and the number of sub-wires. And the ground remote workstation drives the robot to move through 5G communication and carries out X-ray detection to obtain a high-resolution image. Based on the normalized matching degree of the X-ray image of the to-be-detected crimping fitting and the corresponding structural feature template, determining an adaptive denoising threshold value of wavelet transform, denoising the X-ray image, and segmenting the X-ray image into a steel core area and an aluminum pipe area through gray histogram analysis; and carrying out local adaptive enhancement by adopting a multi-scale Retinex algorithm, and obtaining a nondestructive testing result of the power transmission line through a multi-task defect identification network fused by attitude correction and a geometric attention mechanism. The nondestructive testing precision of the power transmission line is effectively improved.

Description

Technical Field

[0001] This invention relates to the field of power transmission line testing technology, and in particular to a non-destructive testing method and system for power transmission lines. Background Technology

[0002] Crimping fittings for overhead transmission lines, such as tension clamps and splicing pipes, are critical components ensuring conductor and ground wire fixation and power transmission. Their crimping quality directly affects the reliability and safety of the transmission line. Because crimping quality issues are hidden inside the fittings, traditional inspection methods primarily rely on manual operation of X-ray flaw detection equipment on the tower, visually assessing defects. This method is not only inefficient, taking an average of about 50 minutes to inspect each fitting and requiring at least five people to work together, but also carries safety risks such as falls from heights, X-ray radiation, and electric shock. Furthermore, manual inspection is limited by the skill level and operational standards of the construction personnel, making it difficult to guarantee consistency and accuracy. This is especially true in complex line structures or harsh environments, further increasing the difficulty of inspection and leading to a high rate of missed crimping defects; for example, the defect rate of tension clamps in three-span sections is as high as 29.09%.

[0003] In recent years, some regions have experimented with drone-based X-ray inspection technology, directly mounting flaw detection equipment on drones to reduce manual high-altitude operations. Other research has focused on using drone platforms equipped with motion-adjustable inspection robots to achieve precise flaw detection of press-fit fittings. In this method, the drone, with its agility and maneuverability, carries the robot to the target fitting location. Through visual positioning and attitude adjustment technology, the robot's X-ray flaw detection module maintains the optimal inspection distance and angle with the fitting. The robot then activates its flaw detection equipment to acquire images of the fitting's internal structure, providing data support for subsequent defect identification. In the image preprocessing stage, to reduce noise interference in defect assessment, wavelet transform technology is commonly used for X-ray image denoising. This technology decomposes the image signal into wavelet coefficients of different scales, suppressing high-frequency, small coefficients representing noise while retaining low-frequency, large coefficients representing image details. This achieves the dual goals of noise removal and image clarity, laying the foundation for defect identification.

[0004] While drone-based robot inspection methods have significantly improved operational safety and efficiency, wavelet transform-based X-ray image denoising technology still has unresolved shortcomings. The core problem with the widely used standard wavelet transform technology lies in its use of fixed wavelet basis functions and a single decomposition scale for signal processing, which fails to dynamically adapt to the complex distribution characteristics of noise and defect features in X-ray images. Noise in X-ray images often presents as random high-frequency signals, and critical defects such as micro-cracks and under-pressure edges inside crimp fittings also exhibit characteristic signals in the high-frequency band, with signal strengths often similar to noise. Standard wavelet transform lacks an effective differentiation mechanism when identifying these high-frequency signals, easily misjudging these high-frequency defect features as noise and over-smoothing them, leading to weakening or even complete elimination of defect features. This technical deficiency directly results in the missed detection of hidden dangers inside crimp fittings, significantly reducing the accuracy of drone-based robot inspection methods, hindering the full realization of their technical advantages, and failing to provide reliable quality assurance for the inspection of crimp fittings in power transmission lines. Summary of the Invention

[0005] Therefore, the technical problem to be solved by the present invention is to overcome the shortcomings of the existing detection method of UAV-borne robots, which uses fixed wavelet basis functions and a single decomposition scale for signal processing, and cannot dynamically adapt to the complex distribution characteristics of noise and defect features in X-ray images, resulting in poor accuracy of non-destructive testing of transmission lines.

[0006] To solve the above-mentioned technical problems, the present invention provides a non-destructive testing method for transmission lines, comprising: For X-ray images of defect-free steel-cored aluminum stranded wires corresponding to different types of crimping fittings, the frequency, direction and grayscale features of periodic winding textures are extracted to generate steel-cored aluminum stranded wire texture maps for each type of crimping fitting. Based on the spatial position and size parameters of each component of each type of crimping fitting, feature point matching and spatial registration are performed on the corresponding defect-free steel-cored aluminum stranded wire X-ray images to generate the fitting structure position mask for each type of crimping fitting. By channel fusion of the steel-cored aluminum stranded wire texture map and the structural position mask of each type of crimping fitting, a structural feature template of each type of crimping fitting is obtained; Acquire X-ray images of the crimping fitting to be inspected and match them with their corresponding structural feature templates; obtain the normalized matching degree between the two and determine the adaptive denoising threshold of wavelet transform based on the normalized matching degree; perform wavelet transform on the X-ray images of the crimping fitting to be inspected to obtain the denoised X-ray images. Based on the target X-ray image, the non-destructive testing results of the crimp fitting to be tested are obtained.

[0007] Preferably, the process of obtaining the normalized matching degree between the X-ray image of the crimp fitting to be inspected and the corresponding structural feature template includes: The X-ray image to be detected is divided into multiple sub-windows. The wavelet coefficient distribution in each sub-window is extracted, and the normalized cross-correlation coefficient between the wavelet coefficient and the sub-window in the same coordinate range in the matching structural feature template is calculated. This cross-correlation coefficient is then used as the normalized matching degree of the sub-window.

[0008] Preferably, the formula for determining the adaptive denoising threshold of the wavelet transform for each sub-window based on the normalized matching degree of each sub-window is as follows: , in, The first X-ray image of the crimping fitting to be inspected Line number Adaptive denoising threshold for sub-windows of a column. As the global base threshold, As a regulating factor, The first X-ray image of the crimping fitting to be inspected Line number The normalized cross-correlation coefficient between the sub-windows of the column and the sub-windows within the same coordinate range in the matched structural feature template.

[0009] Preferably, the method for acquiring X-ray images of the crimp fitting to be inspected includes: Using drones with a dedicated mounting device, an airborne non-destructive testing robot is precisely deployed to the location of the crimp fittings to be tested on the power transmission line; Using a ground-based remote workstation, an airborne non-destructive testing robot is driven to move along the power transmission line, and X-ray equipment is used to collect X-ray images of the crimped fittings to be tested.

[0010] Preferably, the method for obtaining the non-destructive testing results of the crimp fitting to be tested based on the target X-ray image includes: Based on the principal edge features of the conductor in the current target X-ray image coordinate system, the rotation angle of the conductor in the image plane is obtained. Based on the rotation angle of the conductor in the image plane and the tilt angle of the robot body, an affine transformation is performed on the target X-ray image to obtain a geometrically corrected target X-ray image. The geometrically corrected X-ray image of the target is used to obtain the non-destructive testing results of the press-fit fitting to be tested through a non-destructive testing model.

[0011] Preferably, the non-destructive testing model is a deep convolutional neural network, which includes: a backbone network, a geometric attention mechanism module, and a classifier.

[0012] Preferably, the method for obtaining the non-destructive testing results of the crimped fitting by passing the geometrically corrected target X-ray image through a non-destructive testing model includes: The geometrically corrected target X-ray image is passed through the backbone network to extract deep feature maps; The structural feature template corresponding to the X-ray image of the press-fit fitting to be inspected is used to generate a geometric attention mask through the geometric attention mechanism module; The deep feature map is fused with the geometric attention mask to obtain the target feature map; The target feature map is passed through a classifier to obtain the non-destructive testing results of the crimp fitting to be tested.

[0013] Preferably, the method for obtaining the non-destructive testing results of the crimp fitting to be tested based on the target X-ray image further includes: Gray-level histogram analysis was performed on the target X-ray image to segment it into a steel core region and an aluminum tube region; For different regions obtained by segmentation, the multi-scale Retinex algorithm is used for local adaptive enhancement to obtain enhanced target X-ray images; among them, the Gaussian wrapping scale parameter and gain coefficient of the steel core region are greater than those of the aluminum tube region. Based on the enhanced X-ray image of the target, the non-destructive testing results of the crimp fitting to be tested are obtained.

[0014] Preferably, the non-destructive testing results of the crimp fitting to be tested include: defect type, defect location coordinates, defect size, and confidence level report.

[0015] The present invention also provides a non-destructive testing system for transmission lines, comprising: The texture map construction module is used to extract the frequency, direction and grayscale features of the periodic winding texture from X-ray images of defect-free steel-cored aluminum stranded wires corresponding to different types of crimping fittings, and generate texture maps of steel-cored aluminum stranded wires for each type of crimping fitting. The position mask generation module is used to perform feature point matching and spatial registration on the corresponding defect-free steel-cored aluminum stranded wire X-ray images based on the spatial position and size parameters of each component of each type of crimping hardware, and generate the hardware structure position mask of each type of crimping hardware. The structural feature template construction module is used to perform channel fusion of the steel core aluminum stranded wire texture map and the structural position mask of each type of crimping hardware to obtain the structural feature template of each type of crimping hardware; The denoising module is used to acquire X-ray images of the crimping fitting to be inspected, match them with their corresponding structural feature templates, obtain the normalized matching degree between the two, determine the adaptive denoising threshold of wavelet transform based on the normalized matching degree, and perform wavelet transform on the X-ray image of the crimping fitting to be inspected to obtain the denoised X-ray image. The inspection module is used to obtain the non-destructive testing results of the crimped fittings to be inspected based on the target X-ray image.

[0016] Compared with the prior art, the above-described technical solution of the present invention has the following advantages: This invention discloses a non-destructive testing method and system for power transmission lines. Based on defect-free X-ray images of different types of crimp fittings, texture features are extracted to generate texture maps. Combined with spatial parameters of the fitting components, image registration is completed to generate a position mask. The two are then fused to obtain a model-specific structural feature template, achieving accurate calibration of high-frequency features of normal fitting structures. Subsequently, the denoising threshold is dynamically adjusted by calculating the normalized matching degree between the image to be tested and the corresponding template. High-threshold filtering is used to remove high-frequency texture signals from normal structural areas that highly match the template, while low-threshold retention is used for potential defect areas that do not match the template. This effectively avoids misjudgment and over-smoothing of high-frequency defect features such as micro-cracks and under-pressure edges by traditional wavelet transform, significantly improving the retention rate of defect features during X-ray image denoising. This effectively improves the accuracy of internal defect detection in crimp fittings, fully leveraging the safety and efficiency advantages of UAV inspection, and providing a reliable guarantee for the quality inspection of power transmission line crimp fittings.

[0017] To address the issue that traditional multi-scale Retinex algorithms are prone to producing "halo" artifacts at material boundaries when enhancing images of high-density components such as steel cores and aluminum tubes, this paper proposes a new approach. Furthermore, the algorithm may indiscriminately over-enhance random noise in the background region, interfering with subsequent accurate defect identification. This invention uses grayscale histogram analysis to precisely segment images into steel core and aluminum tube regions. Different Gaussian wraparound scale parameters and gain coefficients are then applied to the two regions based on their different characteristics. For the steel core region, where grayscale values ​​are higher and defect details are easily obscured, a larger Gaussian wraparound scale parameter and gain coefficient are used to enhance the grayscale difference between deep defects and the background, thus highlighting minute defects. For the aluminum tube region, where grayscale distribution is relatively uniform, a smaller Gaussian wraparound scale parameter and gain coefficient are used to moderately improve clarity while avoiding excessive enhancement of boundaries. This method abandons the enhancement mode of traditional algorithms that use globally uniform parameters. It suppresses halo artifacts at the steel-aluminum material boundary and avoids excessive amplification of random noise in the background region, improving the effective feature contrast of X-ray images. This lays a high-quality image foundation for subsequent accurate defect identification and further ensures the accuracy and reliability of non-destructive testing of transmission line crimping fittings.

[0018] To address the issue that the highly similar features between crimping defects (such as leaks) and normal crimping deformations can easily lead to confusion and misjudgment by deep convolutional neural networks (DCNNs), and the angular deviations often present in images acquired on-site due to drone movement or robot posture changes, further complicate DCNN recognition, this invention employs an affine transformation on the image before it enters the backbone network for analysis. This transformation corrects the image to a standardized "frontal view," reducing feature differences caused by changes in shooting angle and significantly improving the network's recognition accuracy and robustness in complex on-site environments. Furthermore, an attention weight map is introduced in the deep feature processing stage to guide computational resources, dynamically generating a geometric attention mask that matches the size of the deep feature map. In "defect-prone areas" (such as steel anchors, crimping areas, and anti-slip grooves), the mask is assigned high weight values. In non-critical or defect-infrequent areas, the weight values ​​are significantly reduced. In this way, the network's characteristic responses to high-weight regions are enhanced and preserved, while the characteristic responses to low-weight regions are suppressed or weakened, effectively improving the accuracy of non-destructive testing of transmission line crimp fittings. Attached Figure Description

[0019] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings, wherein:

[0020] Figure 1 This is a flowchart of the steps of a non-destructive testing method for power transmission lines according to the present invention.

[0021] Figure 2 This is a flowchart illustrating a non-destructive testing method for power transmission lines according to the present invention.

[0022] Figure 3 This is a schematic diagram of a drone carrying an airborne non-destructive testing robot via a dedicated mounting device. Detailed Implementation

[0023] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention.

[0024] Reference Figure 1 , 2 As shown, this embodiment provides a non-destructive testing method for transmission lines, including: Step S1: Extract the frequency, direction and grayscale features of the periodic winding texture from the X-ray images of the defect-free steel-cored aluminum stranded wire corresponding to different models of crimping fittings, and generate the steel-cored aluminum stranded wire texture map of each model of crimping fitting. In this embodiment, specifically, a clear X-ray image of a defect-free steel-cored aluminum stranded wire corresponding to the target model crimping fitting is selected as a standard sample; the background area of ​​the image is removed, and the effective area of ​​the steel-cored aluminum stranded wire is retained; and it is rotated and corrected to a horizontal or vertical reference position to ensure that the texture analysis reference is consistent.

[0025] Steel-cored aluminum stranded wire is made of multiple metal wires spirally wound together, which appear as alternating light and dark stripes under X-ray fluoroscopy. The frequency characteristic of the periodic winding texture is the fixed spacing between two stripes, much like recognizing the beat of music. By identifying the speed of this recurring "light-dark-light-dark" rhythm, the system can determine the density of the wire's winding, i.e., the frequency characteristic of the periodic winding texture.

[0026] The metal wires in a steel-cored aluminum stranded wire are not laid flat horizontally or vertically, but rather distributed at a specific tilt angle. Multi-angle scanning is performed on X-ray images of defect-free steel-cored aluminum stranded wires. The detection signal intensity reaches its maximum when the scanning direction perfectly aligns with the tilt angle of the metal wires. In this way, the system accurately records the tilt direction (left or right) and specific tilt angle of the metal wires; these recorded tilt direction and tilt angle parameters constitute the directional characteristics.

[0027] Analyzing texture grayscale features, or extracting grayscale features, involves examining the grayscale depth of X-ray images to determine the thickness and density of an object. Steel-cored aluminum stranded wire has a cylindrical structure, exhibiting a shape difference where it is thicker in the middle and thinner at the sides, and the steel core and aluminum wire have different densities. The system analyzes the variation patterns of grayscale (brightness) in the image, identifying areas with the lowest grayscale values ​​corresponding to the areas of maximum metal thickness or density, and areas with gradually increasing grayscale values ​​corresponding to areas of gradually decreasing metal thickness. By statistically analyzing the distribution of grayscale (brightness), the normal grayscale distribution range of defect-free steel-cored aluminum stranded wire in X-ray images is determined; this normal grayscale distribution range is the grayscale feature.

[0028] The three elements extracted above—stripes density (frequency), tilt angle (direction), and material brightness (grayscale)—are recombinated. A computer-generated image of a perfect conductor under "ideal conditions," free of noise and defects, is then created—a texture map. This texture map serves as the ideal texture benchmark for the steel-cored aluminum stranded wire corresponding to the target model of the crimping hardware. It can be used as a benchmark for defect detection. During comparison, conductors that perfectly match the grayscale distribution, texture period, and direction of the texture map are considered normal conductors; those that do not match are deemed to have defects, such as extra spots or broken lines.

[0029] Step S2: Based on the spatial position and size parameters of each component of each type of crimping fitting, feature point matching and spatial registration are performed on the corresponding defect-free steel-cored aluminum stranded wire X-ray images to generate the fitting structure position mask for each type of crimping fitting. In this embodiment, specifically, standardized data for all models of crimping fittings (such as tension clamps and splicing tubes) are collected. The structures of different models of crimping fittings are different, and corresponding CAD design drawings and three-dimensional solid models need to be obtained according to the model. The precise spatial coordinates, dimensional parameters and relative positional relationships of key structures such as "steel anchor", "anti-slip groove", "crimping area", "non-crimping area" and "steel core channel" of each model of fitting are clearly marked.

[0030] Collect standard image samples of defect-free fittings: Classify fittings by model and collect standardized images separately for each model of qualified crimping fittings to ensure that the samples cover the steel-cored aluminum stranded wire winding specifications corresponding to each model of fitting. Finally, archive the samples by model to form a defect-free image dataset. Perform texture analysis on the steel-cored aluminum stranded wire area in the defect-free sample images, such as using Fourier transform or Gabor filtering to extract the frequency, direction and grayscale features of the dominant periodic winding texture, and generate steel-cored aluminum stranded wire texture maps for each model of crimping fittings.

[0031] Classified by hardware model, and based on the corresponding model's CAD drawings or standard images, feature point matching and spatial registration are performed on the standard images of that model. Key structural areas such as "anti-slip groove," "steel anchor," and "pressing area" of the hardware model are digitally marked (e.g., the anti-slip groove area of ​​a certain model is marked as value A, the steel core area is marked as value B, and the aluminum tube area is marked as value C), generating a unique binary or multi-valued hardware structure position mask for that model.

[0032] Step S3: Channel fusion of the steel core aluminum stranded wire texture map and the structural position mask of each type of crimping fitting to obtain the structural feature template of each type of crimping fitting; The structural feature template integrates the hardware structure location mask and the steel-cored aluminum stranded wire texture map to form a multi-channel feature template. This template spatially marks where high-frequency signals should appear (such as the edge of the anti-slip groove and the texture of the steel stranded wire) and where high-frequency signals should not appear (such as the smooth surface of the aluminum tube).

[0033] Step S4: Acquire X-ray images of the crimping fitting to be inspected and match them with their corresponding structural feature templates; obtain the normalized matching degree between the two and determine the adaptive denoising threshold of wavelet transform based on the normalized matching degree; perform wavelet transform on the X-ray images of the crimping fitting to be inspected to obtain denoised X-ray images. In this embodiment, preferably, the process of obtaining the normalized matching degree between the X-ray image of the crimping fitting to be inspected and the corresponding structural feature template includes: The X-ray image to be detected is divided into multiple sub-windows. The wavelet coefficient distribution in each sub-window is extracted, and the normalized cross-correlation coefficient between the wavelet coefficient and the sub-window in the same coordinate range in the matching structural feature template is calculated. This cross-correlation coefficient is then used as the normalized matching degree of the sub-window.

[0034] The formula for calculating the adaptive denoising threshold of wavelet transform is: ,in, The adaptive denoising threshold for wavelet transform. The degree of matching between the current analysis sub-window and the structural feature template. These are wavelet coefficients.

[0035] The formula for determining the adaptive denoising threshold of the wavelet transform for each sub-window based on the normalized matching degree of each sub-window is as follows: , in, The first X-ray image of the crimping fitting to be inspected Line number Adaptive denoising threshold for sub-windows of a column. The global baseline threshold, calculated based on a general denoising algorithm, is used to characterize the overall noise level of the image. This is an adjustment factor, i.e., a preset enhancement coefficient, with a value ranging from 0.5 to 2.0, and a preferred value of 1.0. It is used to control the suppression intensity of structural texture. The first X-ray image of the crimping fitting to be inspected Line number The normalized cross-correlation coefficient between the sub-windows of the column and the sub-windows in the same coordinate range in the matched structural feature template, with a value range of [0,1].

[0036] Based on the above functional relationship, the specific execution logic of the structure-prior-guided wavelet adaptive threshold denoising algorithm is as follows: when When the value approaches 1, it indicates that the current sub-window highly matches the structural feature template, meaning it is identified as a normal structure such as anti-slip groove or steel strand texture. To increase the frequency of texture signals in the area, the algorithm uses a higher threshold to filter out or smooth the high-frequency texture signals in that area, so as to avoid misjudging them as defects. when When the value approaches 0, it indicates that the current sub-window does not match the structural feature template, i.e., it is identified as a potential defect area or anomaly signal. Falling back to The algorithm uses a low base threshold to preserve microscopic edge information and defect details at the location to the maximum extent, preventing fine cracks from being smoothed out.

[0037] If the current sub-window highly matches the structural feature template—for example, the algorithm identifies the sub-window as the edge of an anti-slip groove or the winding texture of a steel core—this indicates that the high-frequency wavelet coefficients here are very likely part of a normal structure. Therefore, the denoising threshold of this sub-window should be increased to filter out these high-frequency signals as "noise" or "normal structure."

[0038] If the current sub-window does not match the structural feature template, for example, if a high-frequency signal appears in an area that should be smooth like an aluminum tube, it indicates that there is a high probability of an abnormal signal. Therefore, the denoising threshold should be lowered (or the coefficient should be retained) to preserve the microscopic edge information of the real defect to the greatest extent possible.

[0039] Existing technologies often employ standard wavelet transform for denoising and the multi-scale Retinex (MSR) algorithm. Research has found that when filtering noise from X-ray images, standard wavelet transform carries the risk of misclassifying high-frequency defect features such as micro-cracks and under-pressure edges inside press-fit fittings as noise and over-smoothing them, which directly leads to missed defects.

[0040] To address the aforementioned issues, this invention proposes a wavelet adaptive threshold denoising algorithm guided by prior structural knowledge. Utilizing prior geometric knowledge of crimped fittings, such as the periodic winding texture of steel-cored aluminum stranded wire and the fixed position of anti-slip grooves, a structural feature template is constructed. The adaptive threshold of the wavelet transform is determined by the matching degree between the image and the structural feature template. During denoising, the algorithm no longer uses a globally uniform static threshold but dynamically adjusts the denoising threshold based on the matching degree between the current analysis region and the template. This ensures that while effectively filtering out random noise, it can retain the microscopic edge information of the true defects to the greatest extent possible. The adjustment of the denoising threshold is positively correlated with or has a specific functional relationship with the matching degree between the image and the structural feature template, aiming to protect "non-matched" high-frequency signals and smooth "matched" high-frequency signals (normal structures).

[0041] Step S5: Based on the target X-ray image, obtain the non-destructive testing results of the crimp fitting to be tested.

[0042] In this embodiment, preferably, the method for obtaining the non-destructive testing results of the crimp fitting to be tested based on the target X-ray image includes: The target X-ray image is standardized to a uniform input viewpoint using a front-end attitude correction module. Based on the principal edge features of the conductor in the current target X-ray image coordinate system, the rotation angle of the conductor in the image plane is obtained. Based on the rotation angle of the conductor in the image plane and the tilt angle of the robot body, an affine transformation is performed on the target X-ray image to obtain a geometrically corrected target X-ray image. The geometrically corrected X-ray image of the target is used to obtain the non-destructive testing results of the press-fit fitting to be tested through a non-destructive testing model.

[0043] This invention utilizes edge detection technology combined with Hough transform to analyze target X-ray images. The goal of this step is to automatically identify and extract the principal edge features of the conductor in the current image coordinate system and calculate the conductor's rotation angle within the image plane. The algorithm integrates physical pose data with the visually extracted rotation angle to calculate the affine transformation matrix required for the image to deviate from a "standard frontal view" (e.g., a zero-degree angle perpendicular to the conductor's axis). This matrix includes rotation, minor scaling, and necessary translation parameters. The calculated affine transformation matrix is ​​applied to the original X-ray image, and through image resampling technology, a geometrically corrected target X-ray image is generated.

[0044] This geometric correction process unifies the target X-ray image to a standardized input viewpoint, greatly eliminating feature changes caused by differences in shooting angles. This allows the core deep convolutional neural network (CNN) to avoid expending resources to learn redundant features caused by angle changes, thus concentrating computational power on the essential feature recognition of defects, improving the network's generalization ability and final recognition accuracy.

[0045] In this embodiment, preferably, the non-destructive testing model is a deep convolutional neural network, which includes: a backbone network, a geometric attention mechanism module, and a classifier.

[0046] Methods for obtaining the non-destructive testing results of the press-fit fitting by passing the geometrically corrected X-ray image of the target through a non-destructive testing model include: The geometrically corrected target X-ray image is passed through a backbone network (such as ResNet-50) to extract deep feature maps; The structural feature template corresponding to the X-ray image of the press-fit fitting to be inspected is used to generate a geometric attention mask through the geometric attention mechanism module; The deep feature map is fused with the geometric attention mask to obtain the target feature map; The target feature map is passed through a classifier to obtain the non-destructive testing results of the crimp fitting to be tested.

[0047] This invention introduces a geometric attention mechanism into the deep feature maps of a deep convolutional neural network. This mechanism is designed based on prior knowledge of defects (e.g., "pressure leakage in the anti-slip groove" must occur at a specific location on the steel anchor), guiding the network to concentrate computational resources on areas with high defect incidence. This makes the network more sensitive to the identification of specific types of defects, while effectively reducing the misjudgment rate of normal deformation in non-critical areas.

[0048] Unlike random image recognition tasks, defects in transmission line crimping fittings (such as leakage or steel core breakage) have a strong location correlation. This invention fully utilizes prior knowledge: for example, "'anti-slip groove leakage' must occur at a specific location on the steel anchor; 'steel core fracture' only occurs in the projected area of ​​the steel core. Certain areas, such as the smooth transition zone or edge attachments of the crimped pipe, although structurally complex, are not typical locations for the target defect." The geometric attention mechanism guides computational resources by introducing an attention weight map during the deep feature processing stage of the network. Based on structural feature templates or precisely registered CAD models, a geometric attention mask matching the size of the deep feature map is dynamically generated. In "high-defect areas" (such as steel anchors, crimped areas, and anti-slip grooves), the mask is assigned high weight values. In non-critical or low-defect areas, the weight values ​​are significantly reduced. During the deep feature extraction process of the CNN, this attention mask is multiplied element-wise with the feature map (or fused through a gating mechanism). In this way, the network's feature responses to high-weight areas are strengthened and preserved, while the feature responses to low-weight areas are suppressed or weakened.

[0049] The standard MSR algorithm is prone to producing "halo" artifacts at material boundaries when enhancing images of high-density components such as steel cores and aluminum tubes. Furthermore, this algorithm may indiscriminately over-enhance random noise in the background, interfering with subsequent accurate defect identification. To address this, this invention designs an adaptive multi-scale Retinex enhancement algorithm based on material grayscale partitioning. This algorithm first automatically segments the image based on the significant grayscale partitioning characteristics of different metal components (steel and aluminum) in X-ray images due to density differences. Then, it applies a multi-scale Retinex algorithm with different parameters to different partitions for local adaptive enhancement. For example, a stronger enhancement is applied to the steel core region with higher grayscale values ​​to highlight deep defects. This method not only effectively improves the visibility of defects in dark areas but also significantly suppresses halo artifacts common in traditional algorithms. The specific scheme is as follows: In this embodiment, preferably, the method for obtaining the non-destructive testing results of the crimp fitting to be tested based on the target X-ray image further includes: Gray-level histogram analysis was performed on the target X-ray image to segment it into a steel core region and an aluminum tube region; For different regions obtained by segmentation, the multi-scale Retinex algorithm is used for local adaptive enhancement to obtain enhanced target X-ray images; among them, the Gaussian wrapping scale parameter and gain coefficient of the steel core region are greater than those of the aluminum tube region. Based on the enhanced X-ray image of the target, the non-destructive testing results of the crimp fitting to be tested are obtained.

[0050] X-ray detection is based on the Beer-Lambert law. Steel (steel core) has a much higher density than aluminum (aluminum stranded wire / tube). Therefore, in X-ray images, steel and aluminum exhibit "significant grayscale zoning characteristics." The steel core region attenuates X-rays more strongly and will display grayscale values ​​distinctly different from the aluminum tube region, potentially being brighter or darker depending on the imaging plate polarity. Because this grayscale difference is "significant," automatic segmentation can be achieved using relatively mature image processing techniques. Analyzing the grayscale histogram of the entire image, it is likely to exhibit bimodal characteristics (one peak representing steel and the other representing aluminum). An automatic thresholding algorithm, such as Otsu's algorithm, can be used to find the optimal grayscale threshold to separate these two peaks. Alternatively, a K-Means clustering algorithm (with K=2 or K=3, including the background) can be used to automatically cluster pixels into "steel regions" and "aluminum regions," where K is the total number of clusters. After segmentation, the algorithm can apply stronger enhancement to the "steel core region" because it is darker and its details are harder to observe, while avoiding halo artifacts at the "steel-aluminum boundary."

[0051] In this embodiment, specifically, the non-destructive testing results of the crimp fitting to be tested include: defect type, defect location coordinates, defect size, and confidence level report.

[0052] Furthermore, this invention upgrades the single defect classification task into a multi-task learning framework encompassing defect classification, localization, and segmentation. By simultaneously optimizing these three related tasks, the network is driven to learn deeper feature representations with greater generalization capabilities. This not only accurately identifies defect types but also precisely outputs quantified location and size reports. This represents a technological leap from simply "identifying defects" to providing accurate "quantitative diagnosis," providing crucial data support for subsequent full lifecycle quality control platforms.

[0053] Existing technologies are only suitable for inspecting single conductors or ground wires, making it difficult to adapt to the full voltage levels from 35kV to 1100kV, single-circuit to multi-circuit line structures, and the full range of inspection needs from single conductors to eight-split conductors. Furthermore, existing UAV mounting devices are simple in design, lacking resistance to electric field interference and adaptability to complex environments. The inspection equipment is also quite heavy, typically exceeding 50kg, limiting the UAV's endurance and operational flexibility. Simultaneously, image analysis still relies on human experience, resulting in low levels of intelligence and an inability to achieve automatic defect identification and full lifecycle management, thus failing to meet the actual needs for efficient, safe, and intelligent inspection of transmission lines.

[0054] In the existing technology, the X-ray inspection method based on UAVs cannot achieve full-scenario adaptation for all voltage levels from 35kV to 1100kV, different line structures and the number of sub-conductors. The inspection efficiency is low and there are safety risks of high-altitude operation. Image analysis relies on human experience and lacks automation and intelligent means, making it difficult to meet the needs of efficient, safe and accurate inspection and full life cycle management of crimp fittings.

[0055] This invention addresses the full-scenario inspection needs of crimp fittings for overhead transmission lines by combining drones and airborne non-destructive testing robots in collaborative operations. This avoids the risks of manual high-altitude work and enables remote ground control throughout the entire process. The specific solution is as follows: like Figure 3 As shown, Figure 3 This is a schematic diagram of a drone carrying an airborne non-destructive testing robot via a dedicated mounting device.

[0056] Using drones with a dedicated mounting device, an airborne non-destructive testing robot can be precisely deployed to the location of the crimping fittings to be tested on 35kV to 1100kV overhead transmission lines. It is compatible with single-circuit, double-circuit, triple-circuit, and quadruple-circuit line structures and the number of sub-conductors such as single conductor, double-split, quadruple-split, six-split, and eight-split. Using a ground-based remote workstation, and based on 5G communication and high anti-interference technology, an airborne non-destructive testing robot is driven to move along the power transmission line. Combined with lightweight high-energy X-ray equipment, it performs multi-angle inspections and acquires high-resolution digital X-ray images. A deep convolutional neural network combined with the multi-scale Retinex algorithm is used to intelligently analyze digital radiographic images, automatically identifying and quantifying defects such as leakage pressure, underpressure, steel core fracture, flash, and multiple pressure defects in unpressurized areas. The inspection images and diagnostic results are uploaded to the full life cycle quality control platform for the fittings, and a visualized image database of "one line, one sub-database" and "multiple images for a single component" is constructed. It supports the connection with the "machine inspection and control" platform of each network and provincial company to realize dynamic tracking of service status and fault early warning. The safe recovery of airborne non-destructive testing robots can be accomplished by using drones and optimizing the recovery mechanism.

[0057] In this embodiment, the UAV adopts a quadcopter design with a maximum takeoff weight of 30kg and is equipped with a high-precision GPS and visual inertial navigation system to ensure positioning accuracy of ±5cm.

[0058] The dedicated mounting device uses a carbon fiber composite frame, with a weight controlled to within 5kg. It is equipped with a standardized quick-plug interface (positioning pin accuracy ±0.5mm) and an electromagnetic adsorption mechanism, and is fixed to the mounting point on the bottom of the drone with bolts. The mounting device has a built-in center of gravity adjustment module, which uses a servo motor to drive the movement of the counterweight, dynamically adjusting the center of gravity deviation to within 2cm. It can achieve precise deployment and retrieval of airborne non-destructive testing robots, ensuring stable flight and operation of the drone in high-altitude strong winds and electric field environments below 500kV.

[0059] Carbon fiber materials combine high strength and lightweight design, reducing drone energy consumption and extending flight time to over 40 minutes; standardized interfaces and positioning pins ensure quick loading and unloading, reducing operation time; the center of gravity adjustment module is optimized through finite element analysis to maintain the drone's stability in complex airflow, improving deployment accuracy to over 98%.

[0060] The dedicated mounting device connects to the UAV control system via a CAN bus, transmitting positioning signals and counterweight adjustment commands in real time. The UAV interacts with the ground remote workstation via a 5G communication module, receiving deployment / retrieval commands. The signal latency is less than 50ms, ensuring precise coordination.

[0061] The airborne non-destructive testing robot includes the following modules: The mobile mechanism is equipped with a multi-wheel walking system, using a high-precision electric telescopic rod and tilt sensor, and an adjustable wheel track (200mm~1000mm) telescopic roller bracket for real-time monitoring of conductor spacing and line inclination. In this embodiment, the mobile mechanism specifically employs a four-wheel walking system with a drive motor torque of 50 N•m. It is equipped with a V-shaped guide groove (opening angle 60°) and a gravity counterweight (center of gravity 10 cm below the roller axis). Combined with lidar and ultrasonic sensors, it adapts to different conductor spacing and line inclination. In conjunction with a high-torque drive motor and intelligent obstacle-crossing algorithm, it supports crossing adjustment plates, guide wire clamps, and line accessories with a climbing height of up to 50 mm, adapting to line inclination variations ranging from 0° to 30°.

[0062] The inspection unit integrates a lightweight X-ray device based on a Marx generator or Tesla pulse transformer, equipped with an adjustable-angle (±45°) robotic arm, a built-in tilt sensor (accuracy ±0.5°) and a multi-size imaging plate adapter (supporting 200mm~600mm). Optimized through pulse shaping and steepening technology, it achieves an instantaneous power of 960MW and a peak output voltage of 510kV, while keeping the weight under 35kg. It features an intelligent focusing algorithm and a multi-size imaging plate adaptation mechanism. Combined with the tilt sensor and gyroscope, it corrects imaging angle deviation to less than 1°, supporting multi-angle imaging to improve crack detection rate.

[0063] The control mechanism, through electromagnetic shielding, filtering circuits and adaptive control algorithms, combined with the fusion technology of lidar, visual sensors and ultrasonic sensors, ensures stable operation in a strong electric field environment; In this embodiment, the control mechanism uses an embedded processor (ARM Cortex-A72), equipped with an electromagnetic shield (60dB shielding effectiveness) and an adaptive filtering circuit, and integrates data from lidar, vision sensors and ultrasonic sensors to run intelligent obstacle crossing and attitude control algorithms.

[0064] The energy supply system uses high-energy-density lithium batteries, combined with dynamic power management, providing a battery life of over 60 minutes. In this embodiment, the energy supply unit is equipped with a 10Ah high-energy-density lithium battery, and the dynamic power management system optimizes the output.

[0065] The communication module, based on 5G and frequency hopping spread spectrum technology, has a command transmission latency of less than 50ms and a video transmission resolution of 1080p.

[0066] In this embodiment, the communication module integrates a 5G communication unit, supports frequency hopping spread spectrum (frequency range 2.4GHz~5.8GHz) and H.265 video encoding, with a video return resolution of 1080p and a packet loss rate of less than 0.1%.

[0067] The mobile mechanism, with its V-shaped guide groove and counterweight design, enhances its anti-tipping and obstacle-crossing capabilities, traversing 50mm climbing heights and adapting to line inclinations from 0° to 30°. The lightweight design of the X-ray equipment improves UAV mounting efficiency, while intelligent focusing and multi-angle imaging increase defect detection rates to 95%. The control mechanism boasts strong resistance to electric field interference, ensuring stable operation in a 500kV environment. A 5G communication module enables low-latency (<50ms) command transmission and high-definition video feedback. The mobile mechanism and the inspection mechanism are connected via a robotic arm hinge, with servo motors controlling angle adjustment and signals transmitted to the control mechanism via an I2C bus. The control mechanism and the communication module are connected via a high-speed serial port, uploading sensor data and images in real time and receiving ground commands. The energy supply mechanism powers each module via shielded cables, dynamically adjusting power distribution to prioritize the X-ray equipment and communication module. It can achieve full-scenario inspection of crimp fittings, adapting to voltage levels from 35kV to 1100kV, single-circuit to multi-circuit line structures, and single-conductor to eight-split conductors, overcoming interference from line accessories to ensure inspection stability and accuracy.

[0068] In this embodiment, the ground remote workstation includes a high-performance industrial computer (equipped with an NVIDIA GPU, with a computing power of 20 TFLOPS), a 5G wireless communication module, a 17-inch high-definition display screen, and a human-machine interface. The interface integrates a route map, robot status monitoring (position, posture, battery level), and detection progress display modules, and supports touch and keyboard operation.

[0069] The ground-based remote workstation connects to the robot's communication module via a 5G network and employs BCH error correction coding and adaptive filtering technology to resist strong electric field interference. Inspection images and sensor data are transmitted to the workstation via an encrypted protocol. The GPU runs a deep convolutional neural network (CNN) for defect identification, and the results are displayed in real-time on the interface.

[0070] This system enables remote control, visual monitoring, and data management of airborne non-destructive testing robots, ensuring that operators can complete the entire inspection process from the ground. 5G communication ensures low latency (<50ms) and high reliability for command and video transmission. The interface intuitively displays the robot's status and inspection images, making operation simple and reducing training costs. High-performance GPUs support real-time image processing and defect analysis, improving inspection efficiency.

[0071] In this embodiment, the quality control platform for the entire lifecycle of fittings is based on a cloud computing architecture and deployed on a server (10TB storage capacity, 100Gbps processing power). It supports 2D / 3D line diagram mapping, integrates an automatic image stitching algorithm (stitching error <1mm) and databases for "one line, one sub-database" and "one component, multiple images". The platform provides standard API interfaces and is compatible with the State Grid and Southern Power Grid's "machine patrol and control" systems.

[0072] The crimping fitting lifecycle quality control platform receives images and diagnostic data uploaded by workstations via a 5G network, employing HTTPS protocol to ensure data security. Image stitching and defect analysis results are stored in a database, with real-time updates to circuit diagrams, supporting remote querying and dynamic monitoring. It enables archiving, dynamic tracking, and fault early warning of inspection images and diagnostic results, supporting full lifecycle management of crimping fittings. An automatic stitching algorithm, combined with the prior geometry of steel-cored aluminum stranded wire, corrects multi-angle imaging deviations, ensuring image integrity; the visualized database supports cross-regional data sharing, and the fault early warning model predicts defect development trends based on historical data, achieving a 90% accuracy rate; the API interface enables seamless integration with existing systems, improving management efficiency.

[0073] In this embodiment, a deep convolutional neural network (CNN) is used, running on a GPU on a ground workstation. The training dataset contains 100,000 annotated defect images (covering issues such as undervoltage, core breakage, and flash). Wavelet transform is used to denoise the images, and a multi-scale Retinex algorithm is employed to enhance the visibility of defects in dark areas. Combined with Hough transform and edge detection technology, internal defects in the press-fit fittings are automatically identified, and a quantitative report of the defect type, location, and size is output. The defect location accuracy reaches 0.5 mm.

[0074] The X-ray equipment transmits images to the control unit via a high-speed data interface (USB 3.0). The control unit then uploads the images to the workstation via a 5G network. The GPU runs intelligent analysis algorithms, and the processing results are fed back to the interface and management platform. This automatically identifies and quantifies internal defects in press-fit fittings, improving detection accuracy and efficiency, and replacing manual judgment. A CNN algorithm combined with multi-scale Retinex enhances the visibility of defects in dark areas, achieving a 95% identification accuracy and a 0.5mm positioning accuracy. The quantification report provides defect size and location data, supporting precise maintenance decisions. Compared to manual analysis, the processing speed is increased by 10 times, and the false positive rate is reduced.

[0075] This method achieves full-scene crimping fitting inspection through the collaborative operation of a drone and an airborne non-destructive testing robot. The drone precisely deploys the robot onto the power transmission line using a dedicated mounting device. The mounting device's quick-connect interface and center-of-gravity adjustment module ensure stable deployment. The robot's movement mechanism uses a V-shaped guide groove and counterweight to connect to the power line, moving along the line and traversing line edges using high-precision sensors and intelligent obstacle-crossing algorithms. The inspection mechanism employs lightweight X-ray equipment combined with tilt sensors for multi-angle imaging, generating high-resolution digital images. The ground workstation sends control commands via a 5G network to drive the robot's movement and inspection, and receives image and sensor data. The intelligent analysis system automatically identifies defects based on a CNN algorithm and outputs quantitative reports. The inspection results are uploaded to a full lifecycle quality control platform, forming a visualized database that supports dynamic tracking and fault early warning. After inspection, the drone retrieves the robot via the mounting device. The entire process eliminates the need for manual tower climbing, increasing inspection efficiency by 6 times and significantly improving safety and intelligence.

[0076] The signal transmission process is as follows: The ground workstation sends the deployment command to the drone via the 5G network. The drone control system drives the mounting device via the CAN bus. The positioning pin and the electromagnetic adsorption mechanism ensure accurate robot deployment. During retrieval, the reverse operation is performed to complete the docking.

[0077] The workstation sends motion and detection commands, which are transmitted to the robot communication module via the 5G network. The control mechanism distributes commands to the motion and detection mechanism via the I2C bus, and sensor data (LiDAR, vision, ultrasound) is fed back to the workstation via the serial port.

[0078] The X-ray equipment generates images, which are transmitted to the control unit via USB 3.0, uploaded to the workstation via 5G network, and the GPU runs a CNN algorithm to analyze defects. The results are displayed on the interface and uploaded to the management platform.

[0079] The control platform receives images and diagnostic data via HTTPS protocol, automatically stitches together images to correct deviations, stores them in a database, and supports real-time querying and early warning.

[0080] This embodiment solves the technical challenges of full-scene detection, strong electric field interference, and manual reliance through the precise design of each component and efficient signal transmission, and realizes efficient, safe, and intelligent non-destructive testing of power transmission lines.

[0081] The non-destructive testing method for power transmission lines based on UAV collaboration in this embodiment relies on the underlying technical principles of multidisciplinary integration, covering UAV and robot collaborative control, X-ray digital imaging, artificial intelligence image analysis, and high anti-interference communication technology. It aims to achieve full-scenario, efficient, and safe testing of crimping fittings for overhead power transmission lines.

[0082] The collaborative operation of UAVs and airborne non-destructive testing robots is based on Newtonian mechanics, aerodynamics, and motion control theory. The UAV, using a quadcopter dynamics model, precisely adjusts rotor lift and torque using Newton's second law (F=ma) to achieve high-precision positioning (error ±5cm) and stable hovering. A dedicated mounting device employs finite element static analysis to optimize the center of gravity distribution, combined with electromagnetic adsorption and quick-connect interfaces to ensure the stability and accuracy of robot deployment and retrieval. The robot's movement mechanism, through a friction and gravity counterweight design based on static and dynamic principles, reduces the risk of tipping over. A V-shaped guide groove utilizes geometric guidance principles to achieve precise docking with the wire. The control system uses a Kalman filter algorithm to fuse data from lidar, vision sensors, and ultrasonic sensors to correct the motion trajectory in real time, adapting to line inclination (0°~30°) and accessory interference (such as adjustment plates and guide wire clamps).

[0083] X-ray detection is based on Beer-Lambert's law ( , The intensity of the incident X-rays, The intensity of the emitted X-rays, The linear attenuation coefficient is... For X-rays, the thickness of the object along the path of penetration. The X-ray attenuation characteristic, a natural constant, is utilized to generate high-resolution grayscale images by leveraging the attenuation properties of X-rays as they penetrate objects due to differences in material density and thickness. The lightweight high-energy X-ray equipment employs Marx generator technology, generating nanosecond pulses exceeding 510kV through parallel charging and series discharging of multiple capacitors, achieving an instantaneous power of 960MW while keeping the weight under 35kg, meeting the requirements for UAV payload. The imaging plate directly converts X-rays into electrical signals using semiconductor materials (such as silicon), and combines this with a tilt sensor to correct the imaging angle (error <1°), enabling multi-angle imaging and enhancing the visualization of latent defects such as cracks. Image preprocessing employs wavelet transform denoising and multi-scale Retinex algorithms to enhance the visibility of defects in dark areas, ensuring accurate defect identification.

[0084] Defect identification is based on a deep convolutional neural network (CNN). It is trained on a dataset containing 100,000 labeled defect images to extract multi-level features (such as edges, texture, and gray-level gradients). A multi-scale pyramid enhancement algorithm, based on the Laplacian pyramid decomposition principle, reconstructs the image gray-level distribution, highlighting defect features such as undervoltage, leakage, and steel core fracture. Hough transform and edge detection techniques further locate the defect geometry, achieving a defect localization accuracy of 0.5 mm. The algorithm optimizes weights through backpropagation and combines transfer learning to improve its generalization ability to complex defects, achieving an identification accuracy of over 95%, significantly outperforming traditional human experience-based judgment.

[0085] The remote control system adopts H.265 video compression coding and multi-antenna diversity technology, combined with BCH error correction coding and adaptive filtering algorithm, to ensure that the video transmission packet loss rate is less than 0.1% in a 500kV strong electric field environment. It supports real-time visual monitoring and operation interface integration with line map, robot status and detection progress.

[0086] Remote control and video transmission utilize 5G communication combined with frequency hopping spread spectrum (FH) and adaptive filtering technologies. Based on Shannon's information theory and electromagnetic compatibility principles, it ensures low-latency (<50ms) signal transmission even in a 500kV strong electric field environment. Frequency hopping spread spectrum avoids interference by rapidly changing the carrier frequency, adaptive filtering dynamically adjusts parameters to suppress electromagnetic noise, and BCH error correction coding adds redundant check bits to ensure reliable transmission of commands and video data (1080p resolution) with a packet loss rate of less than 0.1%.

[0087] This invention utilizes a drone to deploy an airborne non-destructive testing robot to 35kV–1100kV power lines via a dedicated mounting device, adaptable to various line structures and the number of sub-conductors. A ground-based remote workstation drives the robot's movement and performs X-ray inspection via 5G communication, acquiring high-resolution images. A deep convolutional neural network is employed to automatically identify defects with an accuracy exceeding 95%. Inspection results are uploaded to a full lifecycle quality control platform, constructing a visualized image database that supports dynamic tracking and fault early warning. The drone recovers the robot, increasing inspection efficiency by six times, reducing manpower by half, avoiding the risks of high-altitude operations, and achieving efficient, safe, and intelligent inspection. The dedicated mounting device and lightweight X-ray equipment (35kg) ensure stable operation of the drone in complex electric field and airflow environments, enhancing the safety and reliability of the inspection.

[0088] This invention utilizes a deep convolutional neural network (CNN) combined with a multi-scale pyramid enhancement algorithm to automatically identify internal defects in press-fit fittings (such as under-pressure, leakage, and core breakage), achieving an identification accuracy of 95% and a positioning accuracy of 0.5mm. This overcomes the subjectivity and inefficiency of manual judgment (traditional methods have a false negative rate as high as 29.09%). Multi-angle imaging and image stitching technology ensure comprehensive detection of complex defects, significantly improving detection accuracy. Detection results are uploaded in real-time to a full lifecycle quality control platform, constructing a visualized database of "one line, one sub-database" and "multiple images per component," supporting dynamic tracking and fault early warning, and providing accurate data for operation and maintenance decisions.

[0089] This invention utilizes highly interference-resistant 5G communication (latency <50ms, 1080p video transmission) and intelligent obstacle-crossing algorithms to adapt to complex environments such as mountainous, high-altitude, and coastal areas, solving the problems of traditional technologies being limited to single-conductor detection and lacking applicability across all scenarios. The platform connects with the State Grid and Southern Power Grid's "machine inspection and control" systems to achieve cross-regional data sharing, promoting the digital and intelligent transformation of the detection industry. In engineering construction, operational inspections, and fault diagnosis, it prevents the risk of line breaks in "three-cross" sections, reduces economic losses, and improves power grid reliability, possessing broad industrial application value.

[0090] This invention achieves highly efficient and accurate inspection of crimped fittings across all scenarios by leveraging the dynamic coordination of drones and robots, the physical penetration characteristics of X-rays, the deep learning capabilities of artificial intelligence, and highly interference-resistant communication technology. The collaborative control of drones and robots ensures precise deployment and stable operation, X-ray imaging provides high-resolution defect images, artificial intelligence algorithms enable automated defect identification, and highly interference-resistant communication guarantees real-time remote control. Together, these technologies solve the problems of low efficiency, poor safety, and insufficient intelligence in traditional inspection methods.

[0091] This second embodiment provides a non-destructive testing system for power transmission lines, including: The texture map construction module is used to extract the frequency, direction and grayscale features of the periodic winding texture from X-ray images of defect-free steel-cored aluminum stranded wires corresponding to different types of crimping fittings, and generate texture maps of steel-cored aluminum stranded wires for each type of crimping fitting. The position mask generation module is used to perform feature point matching and spatial registration on the corresponding defect-free steel-cored aluminum stranded wire X-ray images based on the spatial position and size parameters of each component of each type of crimping hardware, and generate the hardware structure position mask of each type of crimping hardware. The structural feature template construction module is used to perform channel fusion of the steel core aluminum stranded wire texture map and the structural position mask of each type of crimping hardware to obtain the structural feature template of each type of crimping hardware; The denoising module is used to acquire X-ray images of the crimping fitting to be inspected, match them with their corresponding structural feature templates, obtain the normalized matching degree between the two, determine the adaptive denoising threshold of wavelet transform based on the normalized matching degree, and perform wavelet transform on the X-ray image of the crimping fitting to be inspected to obtain the denoised X-ray image. The inspection module is used to obtain the non-destructive testing results of the crimped fittings to be inspected based on the target X-ray image.

[0092] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0093] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0094] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0095] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0096] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A non-destructive testing method for transmission lines, characterized in that, include: For X-ray images of defect-free steel-cored aluminum stranded wires corresponding to different types of crimping fittings, the frequency, direction and grayscale features of periodic winding textures are extracted to generate steel-cored aluminum stranded wire texture maps for each type of crimping fitting. Based on the spatial position and size parameters of each component of each type of crimping fitting, feature point matching and spatial registration are performed on the corresponding defect-free steel-cored aluminum stranded wire X-ray images to generate the fitting structure position mask for each type of crimping fitting. By channel fusion of the steel-cored aluminum stranded wire texture map and the structural position mask of each type of crimping fitting, a structural feature template of each type of crimping fitting is obtained; Acquire X-ray images of the crimping fitting to be inspected and match them with their corresponding structural feature templates; obtain the normalized matching degree between the two and determine the adaptive denoising threshold of wavelet transform based on the normalized matching degree; perform wavelet transform on the X-ray images of the crimping fitting to be inspected to obtain the denoised X-ray images. Based on the target X-ray image, the non-destructive testing results of the crimp fitting to be tested are obtained.

2. The non-destructive testing method for transmission lines according to claim 1, characterized in that, The process of obtaining the normalized matching degree between the X-ray image of the crimp fitting to be inspected and the corresponding structural feature template includes: The X-ray image to be detected is divided into multiple sub-windows. The wavelet coefficient distribution in each sub-window is extracted, and the normalized cross-correlation coefficient between the wavelet coefficient and the sub-window in the same coordinate range in the matching structural feature template is calculated. This cross-correlation coefficient is then used as the normalized matching degree of the sub-window.

3. The non-destructive testing method for transmission lines according to claim 2, characterized in that, The formula for determining the adaptive denoising threshold of the wavelet transform for each sub-window based on the normalized matching degree of each sub-window is as follows: ， in, The first X-ray image of the crimping fitting to be inspected Line 1 Adaptive denoising threshold for sub-windows of a column. As the global baseline threshold, As a regulating factor, The first X-ray image of the crimping fitting to be inspected Line 1 The normalized cross-correlation coefficient between the sub-windows of the column and the sub-windows within the same coordinate range in the matched structural feature template.

4. The non-destructive testing method for transmission lines according to claim 1, characterized in that, The method for acquiring X-ray images of the press-fit fitting to be inspected includes: Using drones with a dedicated mounting device, an airborne non-destructive testing robot is precisely deployed to the location of the crimp fittings to be tested on the power transmission line; Using a ground-based remote workstation, an airborne non-destructive testing robot is driven to move along the power transmission line, and X-ray equipment is used to collect X-ray images of the crimped fittings to be tested.

5. The non-destructive testing method for transmission lines according to claim 4, characterized in that, The method for obtaining non-destructive testing results of the crimp fitting to be tested based on the target X-ray image includes: Based on the principal edge features of the conductor in the current target X-ray image coordinate system, the rotation angle of the conductor in the image plane is obtained. Based on the rotation angle of the conductor in the image plane and the tilt angle of the robot body, an affine transformation is performed on the target X-ray image to obtain a geometrically corrected target X-ray image. The geometrically corrected X-ray image of the target is used to obtain the non-destructive testing results of the press-fit fitting to be tested through a non-destructive testing model.

6. The non-destructive testing method for transmission lines according to claim 5, characterized in that, The non-destructive testing model is a deep convolutional neural network, which includes: a backbone network, a geometric attention mechanism module, and a classifier.

7. The non-destructive testing method for transmission lines according to claim 6, characterized in that, Methods for obtaining the non-destructive testing results of the press-fit fitting by passing the geometrically corrected X-ray image of the target through a non-destructive testing model include: The geometrically corrected target X-ray image is passed through the backbone network to extract deep feature maps; The structural feature template corresponding to the X-ray image of the press-fit fitting to be inspected is used to generate a geometric attention mask through the geometric attention mechanism module; The deep feature map is fused with the geometric attention mask to obtain the target feature map; The target feature map is passed through a classifier to obtain the non-destructive testing results of the crimp fitting to be tested.

8. The non-destructive testing method for transmission lines according to claim 1, characterized in that, The method for obtaining non-destructive testing results of the crimp fitting to be tested based on the target X-ray image further includes: Gray-level histogram analysis was performed on the target X-ray image to segment it into a steel core region and an aluminum tube region; For different regions obtained by segmentation, the multi-scale Retinex algorithm is used for local adaptive enhancement to obtain enhanced target X-ray images; among them, the Gaussian wrapping scale parameter and gain coefficient of the steel core region are greater than those of the aluminum tube region. Based on the enhanced X-ray image of the target, the non-destructive testing results of the crimp fitting to be tested are obtained.

9. The non-destructive testing method for transmission lines according to claim 1, characterized in that, The non-destructive testing results of the crimp fittings to be tested include: defect type, defect location coordinates, defect size, and confidence level report.

10. A non-destructive testing system for power transmission lines, characterized in that, include: The texture map construction module is used to extract the frequency, direction and grayscale features of the periodic winding texture from X-ray images of defect-free steel-cored aluminum stranded wires corresponding to different types of crimping fittings, and generate texture maps of steel-cored aluminum stranded wires for each type of crimping fitting. The position mask generation module is used to perform feature point matching and spatial registration on the corresponding defect-free steel-cored aluminum stranded wire X-ray images based on the spatial position and size parameters of each component of each type of crimping hardware, and generate the hardware structure position mask of each type of crimping hardware. The structural feature template construction module is used to perform channel fusion of the steel core aluminum stranded wire texture map and the structural position mask of each type of crimping hardware to obtain the structural feature template of each type of crimping hardware; The denoising module is used to acquire X-ray images of the crimping fitting to be inspected, match them with their corresponding structural feature templates, obtain the normalized matching degree between the two, determine the adaptive denoising threshold of wavelet transform based on the normalized matching degree, and perform wavelet transform on the X-ray image of the crimping fitting to be inspected to obtain the denoised X-ray image. The inspection module is used to obtain the non-destructive testing results of the crimped fittings to be inspected based on the target X-ray image.

Images