Video monitoring identification method fusing satellite and unmanned aerial vehicle images

Abstract

The invention provides a video monitoring identification method fusing satellite and unmanned aerial vehicle images, and the method comprises the steps: employing a deep learning-based image distortion correction model for an image which reaches a target resolution and has a vibration damper position in a power transmission line image with texture details, carrying out the self-adaptive correction of a distortion region through the training of a large number of distortion image samples, and obtaining a vibration damper position image; a corrected shockproof hammer image is obtained; generating a power transmission line inspection report according to the conditions of position offset and abnormal working state of the shockproof hammer, and storing and sending the satellite image and the unmanned aerial vehicle image of the abnormal position and the corresponding health degree evaluation result to related personnel; in the subsequent power transmission line inspection process, the vibration damper spectral feature library and the image distortion correction method are continuously updated, the precision of vibration damper position offset detection and working state evaluation is continuously improved through an incremental learning mode, and a dynamically updated and optimized power transmission line monitoring system is formed.

Description

Technical Field

[0001] The present invention relates to the field of information technology, and in particular to a video surveillance recognition method that integrates satellite and unmanned aerial vehicle (UAV) images. Background Art

[0002] Transmission line shock absorbers are devices used to improve the seismic resistance of power transmission systems. They are primarily used to mitigate the impact of structural vibration caused by external factors such as earthquakes on transmission lines and towers. The design and application of shock absorbers play a crucial role in ensuring the safety and stability of power systems. Once the shock absorber is misaligned, its cushioning and vibration-absorbing functions cannot be effectively fulfilled, resulting in a decrease in the transmission line's ability to withstand vibrations. However, when drones are used to capture transmission lines at close range, the reflective properties of the metal surface of the shock absorber can produce varying degrees of highlights in the image. These highlights can cause localized overexposure in the image, resulting in a loss of texture detail, blurred outlines, and even ghosting in severe cases. This highlight distortion directly impacts subsequent image processing and analysis, making it difficult to accurately measure the shock absorber's position offset. Furthermore, the non-uniformity of the highlight reflections can cause discontinuous grayscale distribution of the shock absorber in the image, leading to drastic grayscale variations and distorted brightness information, complicating image segmentation and feature extraction. At the same time, highlights can interfere with the detection of subtle defects such as corrosion and wear on the surface of the shock absorber, potentially missing some critical defects and affecting the accurate assessment of the shock absorber's working condition. Therefore, effectively suppressing the image distortion caused by metal reflections and improving the image quality of the shock absorber is a key issue that needs to be addressed in drone inspection technology. Summary of the Invention

[0003] The present invention provides a video surveillance recognition method that integrates satellite and UAV images, which mainly includes:

[0004] Obtain satellite and drone images, perform multi-scale fusion of the satellite and drone images using pyramid decomposition, and combine the high-frequency details and low-frequency contour information of the two images through weighted averaging to obtain transmission line images that meet the target resolution and have texture details.

[0005] Analyze normal and offset shock-absorbing hammer image samples in historical data, extract spectral features reflecting position offset, build a shock-absorbing hammer spectral feature library based on the spectral features reflecting position offset, perform spectral matching on the fused image, calculate the spectral angle between the fused image and the samples in the feature library, and determine the precise position of the shock-absorbing hammer based on the spectral angle. If the minimum spectral angle is greater than the preset spectral angle threshold, it is determined that the shock-absorbing hammer has position offset.

[0006] For images of the shock absorber position in transmission line images that meet the target resolution and have texture details, a deep learning-based image distortion correction model is used. By training on a large number of distorted image samples, the distorted area is adaptively corrected to obtain the corrected shock absorber image.

[0007] The texture, shape, and color of the shock-absorbing hammer in the calibrated shock-absorbing hammer image are extracted and combined with the hammer's attribute information to construct a multi-dimensional shock-absorbing hammer health assessment index system. The support vector machine algorithm is used to classify the working status of the shock-absorbing hammer, assess the weight of the impact of the shock-absorbing hammer's working status on the health of the shock-absorbing hammer, and calculate the comprehensive health score of the shock-absorbing hammer. If the score is lower than the preset scoring threshold, the shock-absorbing hammer is judged to have an abnormal risk.

[0008] In the event of position deviation or abnormal working status of the shock absorber, a transmission line inspection report is generated, and satellite and drone images of the abnormal location and the corresponding health assessment results are stored and sent to relevant personnel;

[0009] During subsequent transmission line inspections, the system continuously updates the anti-vibration hammer spectral feature library and image distortion correction method. Through incremental learning, the accuracy of anti-vibration hammer position offset detection and operating status assessment is continuously improved, forming a dynamically updated and optimized transmission line monitoring system.

[0010] Regular video monitoring of satellite remote sensing images and drone inspection images of transmission lines is carried out, and the video monitoring data is batch processed and analyzed to identify shock-absorbing hammer targets in real time in the video, extract their motion trajectory and state change information, and integrate and verify them with static image analysis results. By analyzing the correlation patterns and evolution trends of shock-absorbing hammer failures, the video monitoring identification content is fed back.

[0011] The technical solution provided by the embodiment of the present invention may have the following beneficial effects:

[0012] The present invention discloses a video surveillance and recognition method that integrates satellite and drone imagery. This method acquires high-resolution images of power transmission lines by fusing satellite and drone imagery at multiple scales. A spectral feature library and a spectral angle mapping algorithm are used to precisely locate the position of the shock absorber and detect offsets. A deep learning model is used to correct image distortion, extract the multidimensional features of the shock absorber, and construct a health assessment system. Edge detection and shape recognition are used to assess the state of the shock absorber's connections. The present invention also includes functions such as automatic generation of inspection reports, continuous updating of feature libraries and models, and video surveillance analysis. This method achieves comprehensive monitoring of transmission line shock absorbers, improves detection accuracy and efficiency, and provides a strong guarantee for the safe operation of transmission lines. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1The present invention provides a flow chart of a video surveillance recognition method that integrates satellite and UAV images.

[0014] Figure 2 Schematic diagram of a video surveillance recognition method that integrates satellite and UAV images according to the present invention.

[0015] Figure 3 This is another schematic diagram of a video surveillance recognition method that integrates satellite and UAV images according to the present invention. DETAILED DESCRIPTION

[0016] The technical solutions of the present invention will be clearly and completely described below in conjunction with the embodiments. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0017] like Figure 1-3 In this embodiment, a video surveillance recognition method integrating satellite and drone images may specifically include:

[0018] S101. Obtain satellite images and drone images, perform multi-scale fusion of the satellite images and drone images using pyramid decomposition, combine the high-frequency details and low-frequency contour information of the two images through weighted averaging, and obtain a transmission line image that reaches the target resolution and has texture details.

[0019] Satellite image data of the transmission line area is acquired, geometrically corrected by the nearest neighbor resampling method, and image noise is eliminated by a median filter to obtain first image data; based on the geospatial reference information in the first image data, the image collected by the UAV is geometrically corrected by bilinear interpolation, and noise reduction is performed by a Gaussian filter to obtain second image data; pyramid decomposition is performed on the first and second image data, the Laplace operator is used to extract image contour edge features, and a wavelength threshold is set according to the grayscale value difference to obtain high-frequency and low-frequency components; for the high-frequency components, the local area variance is used as the weight coefficient for fusion, and for the low-frequency components, the local energy is used as the weight coefficient for fusion, and the fused data is reconstructed by an inverse pyramid transform to obtain a target image.

[0020] Specifically, satellite imagery of the power transmission line area was acquired from a remote sensing platform based on the acquisition period. The images were geometrically corrected using the nearest neighbor resampling method, and image noise was removed using a median filter. The geospatial reference information for the area was extracted from the remote sensing data to generate the first set of image data. The drone images were geometrically corrected using bilinear interpolation, denoised using a Gaussian filter, and the satellite and drone images were aligned based on the geospatial reference information to generate the second set of image data. A pyramid decomposer was used to decompose the first set of image data into four layers. The Laplace operator was used to extract edge features from each layer. Within each filter bank, wavelength thresholds were set based on grayscale value differences to separate high-frequency and low-frequency components in the image, generating the third set of image data. The second set of image data was decomposed into four layers using a pyramid hierarchy. Each layer was filtered using a discrete wavelet transform. Texture features were enhanced based on band combinations, and edges were smoothed using a Gaussian function to generate the fourth set of image data. Based on the third and fourth image data sets, the corresponding hierarchical image data were normalized. The local variance of the high-frequency components and the local energy of the low-frequency components were used as weighting coefficients, and a weighted average fusion was performed on the data from the two image sets at different levels. The fused image data was reconstructed using an inverse pyramid transform, and the image contours were enhanced using the Sobel operator. The fused image quality was evaluated using a structural similarity index to obtain the final fused image data. During power line inspections, the nearest neighbor resampling method for satellite images achieves geometric correction by selecting the grayscale value of the nearest pixel around the point to be resampled as the grayscale value of the point. This method is suitable for large-scale line inspections. For example, during a power line inspection, if a pixel with a grayscale value of 240 is found to need to be resampled, the resampled pixel value will be 235 because its nearest neighbor has a grayscale value of 235. To address image noise, a 5×5 sliding window is selected for median filtering. The 25 pixel values ​​within the window are sorted, and the median value is taken as the output pixel value. During UAV image processing, bilinear interpolation is used for geometric correction. The four closest pixels around the point to be resampled are selected and weighted based on the distance between them. For example, when correcting a transmission line tower image, if the grayscale values ​​of the four pixels surrounding a resampled point are 220, 225, 230, and 235, respectively, and the corresponding weight coefficients are 0.4, 0.3, 0.2, and 0.1, the corrected grayscale value of the point is 225.5. During pyramid decomposition, a Gaussian filter is used to downsample the image to generate images of different scales. For a 512×512 transmission line image, after four layers of pyramid decomposition, the image sizes of each layer are 256×256, 128×128, 64×64, and 32×32, respectively.In each image layer, edge features are extracted using the Laplacian operator with the operator template [[0, 1, 0], [1, -4, 1], [0, 1, 0]]. Convolution is then performed on the image to generate an edge feature map. During the image fusion stage, the local variance is used as the weighting factor for high-frequency components. This is calculated by calculating the variance of the pixel values ​​within a 3×3 window. For example, if the grayscale values ​​of nine pixels in a region are 200, 205, 210, 215, 220, 225, 230, 235, and 240, respectively, the mean of this region is 220, the variance is 170.3, and the normalized weighting factor is 0.65. For low-frequency components, the local energy is used as the weighting factor. This is calculated as the average sum of squared pixel values ​​within the window. The final quality assessment uses the structural similarity index, which comprehensively considers the similarity of images in terms of brightness, contrast, and structure. A SSIM value greater than 0.8 indicates good fusion quality. In practical applications, the SSIM value of the fused image is usually between 0.85 and 0.95, which can meet the needs of transmission line inspection.

[0021] S102. Analyze normal and offset shock-absorbing hammer image samples in historical data, extract spectral features reflecting position offset, construct a shock-absorbing hammer spectral feature library based on the spectral features reflecting position offset, perform spectral matching on the fused image, calculate the spectral angle between the fused image and the samples in the feature library, determine the precise position of the shock-absorbing hammer based on the spectral angle, and determine that the shock-absorbing hammer has position offset if the minimum spectral angle is greater than a preset spectral angle threshold.

[0022] An anti-shock hammer image is received, and features of the image are enhanced through RGB channel space conversion. A pre-processed image is obtained by Gaussian filtering, and spectral features of reflectivity, texture, and edge mutation are extracted from the pre-processed image to obtain feature data; the feature data are standardized to eliminate dimensional effects, a feature covariance matrix is ​​calculated through principal component transformation, and the covariance matrix is ​​subjected to eigenvalue decomposition to obtain a position feature matrix; a Gaussian kernel function is used to perform mapping transformation on the position feature matrix, and a feature space of normal position and offset position of the anti-shock hammer is constructed through a support vector machine algorithm, and a spectral feature library of the anti-shock hammer is established; a spectral angle is calculated based on the feature space, and the spatial position coordinates of the anti-shock hammer are fitted using the least squares method to obtain a position offset. If the spectral angle is greater than a preset spectral angle threshold and the position offset is greater than a preset distance threshold, it is determined that the anti-shock hammer has a position offset.

[0023] Specifically, shock-absorbing hammer samples were obtained from an image library based on the sampling time period. Image features were enhanced using a red, green, and blue three-channel color space conversion. Image edges were smoothed using a Gaussian filter. Three spectral features, namely reflectance, texture, and edge abruptness, were extracted from the normal and offset shock-absorbing hammer images to obtain the first set of feature data. This first set of feature data was normalized to eliminate dimensionality effects. The feature covariance matrix was calculated using a principal component transformation. The covariance matrix was subjected to eigenvalue decomposition, and eigenvectors with a cumulative contribution exceeding 90% were selected to construct a shock-absorbing hammer position feature matrix to obtain the second set of feature data. A Gaussian kernel function was used to map the second set of feature data. A support vector machine algorithm was used to construct feature spaces for the normal and offset positions of the shock-absorbing hammer. A shock-absorbing hammer spectral feature library was established based on these feature spaces to obtain the third set of feature data. The input image was subjected to color space conversion and Gaussian filtering preprocessing. Spectral features were extracted using the same method, and position features were obtained using a principal component transformation to obtain the fourth set of feature data. Based on the third and fourth sets of feature data, the spectral angle is calculated using cosine similarity. The spatial coordinates of the shock absorber are fitted using the least squares method, and the distance offset between the center point of the shock absorber and the preset installation location is calculated. If the spectral angle is greater than a preset spectral angle threshold of 0.3 and the position offset is greater than a preset distance threshold of 50 mm, the shock absorber is considered to have position offset. The accuracy of the detection results is verified using a confusion matrix. The installation location of the shock absorber is crucial to line safety in high-voltage transmission lines, and position monitoring is achieved through spectral feature extraction. In the shock absorber image sample, the reflectivity in the normal installation position is typically between 0.6 and 0.8. However, when position offset occurs, the reflectivity drops to 0.3 to 0.5, the texture level drops from the normal 0.8 to around 0.5, and the edge abruptness also drops from 0.9 to below 0.6. Spectral feature extraction uses RGB color space conversion, where the R channel reflects the reflectivity of the shock absorber's metal surface, the G channel represents texture information, and the B channel contains edge features. The image is smoothed by a 5×5 Gaussian filter with a filter kernel parameter σ set to 1.5 to eliminate the influence of noise. The extracted spectral features are normalized and the characteristic covariance matrix is ​​calculated. For example, if the eigenvalues ​​of the covariance matrix of a set of data are [2.5, 1.2, 0.3], the corresponding eigenvectors are [0.8, 0.5, 0.3], [0.5, 0.7, 0.4],

[0024] [0.3, 0.4, 0.8], with a cumulative contribution rate of 92.5%. In the feature space construction, the kernel parameter of the Gaussian kernel function is 0.1, and the support vector machine uses a soft margin parameter C of 10. For the input image, the extracted spectral feature vector is [0.45, 0.52, 0.58]. After principal component transformation, the position feature vector is obtained.

[0025] [0.48,0.55]. The spectral angle θ is calculated by cosine similarity, and the calculation formula is:

[0026] cos(θ) = (A·B) / (|A|·|B|), where A and B are the feature vectors to be matched. For example, the spectral angle between a shock-absorbing hammer image and a normal sample in the feature library is 0.35, which is greater than the preset threshold of 0.3. In spatial position calculation, the least squares method is used to fit the coordinates of the shock-absorbing hammer's center point. Assuming the preset installation coordinates of the shock-absorbing hammer are (100, 200) and the actual detected coordinates are (125, 235), the position offset is 52 mm, exceeding the preset threshold of 50 mm. The detection results were verified using a confusion matrix, with 120 true positives, 8 false positives, 95 true negatives, and 7 false negatives, resulting in a detection accuracy of 93.5%. In practical applications, the dual threshold judgment mechanism for spectral angle and position offset effectively reduces the false positive rate. Position offset is only determined when both indicators exceed the threshold, improving detection reliability.

[0027] S103. For an image of the shock absorber position in a transmission line image that reaches a target resolution and has texture details, a deep learning-based image distortion correction model is used to adaptively correct the distorted area through training with a large number of distorted image samples to obtain a corrected shock absorber image.

[0028] A shock-absorbing hammer distorted image sample is obtained, and noise reduction processing is performed on the distorted image sample using a Gaussian filter. A first set of feature data is obtained using a radial distortion and tangential distortion mapping function. Based on the first set of feature data, image features are extracted using a convolutional neural network, and network parameters are iteratively updated using a back-propagation algorithm to obtain training parameters. Based on the training parameters, the feature map is upsampled using bilinear interpolation, and a corrected image is obtained using an adaptive compensation function. Edge features are extracted from the corrected image, straight line boundaries are detected using a Hough transform, and the least squares method is used to fit the contour curve to obtain the shock-absorbing hammer center coordinates. A geometric transformation matrix is ​​established based on the shock-absorbing hammer center coordinates to correct the distorted area.

[0029] Specifically, samples of shock-absorbing hammer distortion images were obtained from a database based on image acquisition time. The images were denoised using a Gaussian filter and annotated using radial and tangential distortion mapping functions. Three characteristic parameters, namely the degree of distortion, direction of distortion, and region of distortion, were extracted for each image to obtain the first set of image data. A convolutional neural network was used to extract image features from the first set of image data. The network parameters were optimized using a backpropagation algorithm. A loss function was calculated based on the distortion characteristic parameters, and the network weights were iteratively updated to obtain the second set of training parameters. A correction network was constructed based on the second set of training parameters. The feature map was upsampled using bilinear interpolation, and a compensation function was established based on the distortion characteristic parameters. The images were then adaptively corrected to obtain the third set of image data. The third set of image data was then assessed for quality using a structural similarity index. Adaptive threshold segmentation was used to segment the images, and morphological operations were used to fill and smooth small regions within the images to obtain the fourth set of image data. Edge features were extracted from the fourth set of image data. Linear boundaries were detected using a Hough transform. The contour curve was fitted using the least squares method, and the center coordinates of the shock-absorbing hammer were calculated based on the curve equation. A geometric transformation matrix is ​​established based on the center coordinates of the shock absorber. The distorted areas are corrected using a projection transformation, and the correction results are evaluated using the peak signal-to-noise ratio to obtain the corrected image. Distortion correction of the shock absorber in transmission line images involves multiple steps. The first step is extracting distortion feature parameters. Radial distortion primarily manifests as centripetal contraction or expansion of the image, typically described by the polynomial k1×r²+k2×r₄, where r is the distance from the pixel to the image center, and k1 and k2 are the distortion coefficients. When k1=0.05, the image exhibits barrel distortion; when k1=-0.05, pincushion distortion occurs. Tangential distortion is caused by lens mounting deviation. In convolutional neural network training, a 5×5 convolution kernel is used for feature extraction, and a Reinforced Luminance (ReLU) activation function is used. The network consists of four convolutional layers, with feature map channels of 32, 64, 128, and 256, respectively. The loss function uses mean squared error (MSE). The initial learning rate is set to 0.001 and decays by 0.1 every 50 epochs. Training is considered converged when the validation set loss falls below 0.01 and does not decrease for five consecutive epochs. The adaptive compensation of the correction network is based on dynamic adjustment of the distortion parameters. For example, if the radial distortion coefficient k1 of a shock-absorbing hammer image is 0.08, the pixel offset is calculated using the compensation function f(r) = -k1 × r². Taking the image center (320, 240) as an example, a point 100 pixels away from the center has a radial displacement of -0.8 pixels. The calculation of the structural similarity index (SSIM) involves three components: brightness contrast l(x, y), contrast c(x, y), and structure s(x, y), with weights of 0.5, 0.3, and 0.2, respectively. For edge feature extraction, the Hough transform detects lines using an accumulator matrix. The minimum voting threshold is set to 100, and the angular resolution is 1 degree.Least squares fitting of the shock-absorbing hammer contour curve yields the quadratic curve equation y = ax² + bx + c, where the coefficients a, b, and c are obtained by solving the normal equations. The projection transformation matrix H is a 3×3 matrix calculated from the correspondence between the source and target points. A peak signal-to-noise ratio (PSNR) exceeding 35 dB for the correction results indicates good correction results. In practical applications, a set of shock-absorbing hammer images exhibited a radial distortion coefficient of 0.06, and tangential distortion coefficients p1 = 0.002 and p2 = 0.001. After processing with the correction network, the structural similarity index increased from 0.82 to 0.95, the peak signal-to-noise ratio reached 38.5 dB, and the mean square error of the contour curve fitting decreased to 0.8 pixels, achieving accurate distortion correction.

[0030] S104. Extract the texture, shape, and color of the shock-absorbing hammer from the corrected shock-absorbing hammer image, and combine the attribute information of the shock-absorbing hammer to construct a multi-dimensional shock-absorbing hammer health assessment index system. Use a support vector machine algorithm to classify the working status of the shock-absorbing hammer, evaluate the weight of the impact of the working status of the shock-absorbing hammer on the health of the shock-absorbing hammer, and calculate a comprehensive health score of the shock-absorbing hammer. If the score is lower than a preset score threshold, it is determined that the shock-absorbing hammer has an abnormal risk.

[0031] The grayscale co-occurrence matrix and third-order color moment of the shock-absorbing hammer image are obtained, and the first eigenvector is obtained by calculating the contrast, entropy, moment of inertia and surface color distribution; the maximum inter-class variance method is used to calculate the attribute thresholds of corrosion, crack degree and integrity based on the first eigenvector, and the second eigenvector is obtained through the region growing algorithm; the historical data of the shock-absorbing hammer's usage time, quality parameters and vibration frequency are obtained for the second eigenvector, and the eigenvectors with a cumulative contribution rate exceeding a preset threshold are selected through principal component analysis to obtain the third eigenvector; a classifier is constructed using a support vector machine based on the third eigenvector, and the kernel function parameters are optimized through grid search to obtain the health score of the shock-absorbing hammer. If the health score is lower than the preset score threshold, it is determined that the shock-absorbing hammer has an abnormal risk.

[0032] Specifically, based on the acquisition time of the shock-absorbing hammer image, three texture features (contrast, entropy, and moment of inertia) were extracted using a four-directional grayscale co-occurrence matrix. The third-order color moment was used to calculate the surface color distribution characteristics of the shock-absorbing hammer. Ten sets of Fourier descriptor coefficients were extracted for the shape contour. All feature data were normalized to obtain the first set of eigenvectors. For the three image attributes of the shock-absorbing hammer (metal corrosion, surface cracking, and component integrity), the maximum inter-class variance method was used to calculate attribute thresholds. The attribute regions were segmented using a region growing algorithm to obtain the second set of eigenvectors. Historical data (use duration, quality parameters, and vibration frequency) was collected from the shock-absorbing hammer installation records. Principal component analysis was used to calculate eigenvalues ​​and eigenvectors. Eigenvectors with cumulative contributions exceeding 90% were selected to construct a feature matrix, resulting in the third set of eigenvectors. Data from the three sets of eigenvectors were fused, and a four-class classifier was constructed using a support vector machine. The penalty factor and kernel function parameters were optimized through grid search to classify the shock-absorbing hammer's operating status, resulting in the fourth set of eigenvectors. A hierarchical evaluation index was established based on the fourth set of eigenvectors. The weights of the image feature indices were calculated using the entropy weight method, and the weights of the historical data indices were determined using the Delphi method, resulting in the fifth set of eigenvectors. A health score function was constructed for the fifth set of eigenvectors, and a comprehensive score was calculated through weighted summation. If the score falls below a preset threshold of 0.7, the shock absorber is considered to have an abnormal risk, resulting in the shock absorber health assessment result. In the shock absorber health assessment, texture features are extracted using a gray-level co-occurrence matrix. Taking the co-occurrence matrix with a distance of 1 in the 0-degree direction as an example, contrast reflects the intensity of grayscale contrast and is calculated as Σ|ij|2p(i,j), where i and j are the grayscale indexes in the co-occurrence matrix and p(i,j) is the normalized probability distribution of the co-occurrence matrix. The contrast of a normal shock absorber surface is approximately 85, dropping to around 45 when corrosion occurs. The entropy value represents texture complexity and is calculated as -Σp(i,j)log(p(i,j)). A normal value is 4.2, rising to 5.8 when the surface is cracked. In the color moment feature, the first-order moment represents the average color, the second-order moment represents the variance, and the third-order moment represents the skewness. The third-order color moment values ​​of a normal shock-absorbing hammer in the RGB space are

[0033] [120,125,128], [15,18,20], [0.8,0.9,1.0]. After the corrosion, the third-order moment value of the shock-absorbing hammer becomes [85,90,95], [25,28,30], [1.5,1.6,1.8]. The Fourier descriptor is obtained by Fourier transforming the edge contour. The first 10 coefficients are taken to describe the shape characteristics. The coefficient modulus values ​​of the normal shock-absorbing hammer are [100,85,62,45,32,25,18,12,8,5]. In attribute segmentation, the maximum inter-class variance method is used to calculate the corrosion threshold. Assume that the image grayscale range is [0,255] and the image is divided into two classes. The inter-class variance σ 2 =w1w2(μ1-μ2)2 , where w1 and w2 are the ratios of the two types of pixels, and μ1 and μ2 are the grayscale means of the two types. When the threshold corresponding to the maximum inter-class variance is 128, the corrosion area can be accurately segmented. Principal component analysis is used to reduce the dimension of the historical data of the shock-absorbing hammer. The original features include usage time (months), mass loss rate (%), and vibration frequency (Hz). Calculate the eigenvalue [2.5, 0.8, 0.2] and eigenvector

[0034] [[0.8,0.5,0.3],[0.5,0.7,0.4],[0.3,0.4,0.8]], select the first two eigenvectors with a contribution rate of 91%. The support vector machine uses the RBF kernel function and determines the optimal parameters C=10,

[0035] γ = 0.1. The shock-absorbing hammer's status is classified into four categories: normal, slightly abnormal, moderately abnormal, and severely abnormal, with a training accuracy of 92%. The scoring weights are calculated using the entropy weighting method, with image features weighted at 0.6 and historical data weighted at 0.4. A shock-absorbing hammer has a contrast score of 0.65, an entropy score of 0.72, a color moment score of 0.58, a usage time score of 0.8, and a vibration score of 0.75. The weighted calculation yields a comprehensive health score of 0.68, which is below the threshold of 0.7 and indicates an abnormality risk.

[0036] By applying edge detection methods to the corrected shock-absorbing hammer image, the boundaries and contours of the shock-absorbing hammer components are highlighted. Shape recognition methods are used to identify the shapes corresponding to the connected components. The pixel intensity gradients near the shapes corresponding to the connected components are analyzed to accurately locate the connections between components. The comprehensive health score of the shock-absorbing hammer connection is then evaluated. If the score is lower than the preset scoring threshold, it is determined that there is an abnormal risk in the shock-absorbing hammer connection.

[0037] The Sobel operator is used to extract the edge gradient amplitude and direction from the shock-absorbing hammer image, and the edge points are refined using the non-maximum suppression method to obtain the shock-absorbing hammer contour data. Circular components are detected using the circular Huff transform for the shock-absorbing hammer contour data, and the connectors are segmented using the region growing algorithm to obtain the connection region data. The Otsu algorithm is used to calculate the binarization threshold of the connection region data, and the connection edges are enhanced through dilation and erosion operations to obtain the connection feature data. A convolutional neural network is constructed for classification of the connection feature data, and the edge strength score weight is determined using the entropy weight method. If the comprehensive score is lower than the preset score threshold, it is determined that the corresponding shock-absorbing hammer connection has an abnormal risk.

[0038] Specifically, based on the shock-absorbing hammer correction image, the Sobel operator is used to extract the edge gradient amplitude and direction in the horizontal and vertical directions. The non-maximum suppression method is used to refine the edge points. A double threshold is set for the grayscale value of the edge points to perform edge tracking and connection to obtain the first set of contour data. For the first set of contour data, the circular Huff transform is used to detect the circular components of the shock-absorbing hammer. The shock-absorbing hammer connector is segmented using the region growing algorithm. The distance threshold between the connected components is calculated based on the center coordinates and radius, and the connection area is located to obtain the second set of regional data. For the second set of regional data, the Otsu algorithm is used to calculate the local binarization threshold. The connection edges are enhanced through dilation and erosion operations. The connection skeleton line is extracted using the refinement algorithm. The connection contour features are extracted to obtain the third set of feature data. Based on the third set of feature data, the morphological gradient operator is used to calculate the edge strength of the connection. The distance transform is used to calculate the gap distribution of the connection surface. The fourth set of feature data is obtained by curve fitting of the connection edges. A convolutional neural network was constructed based on the fourth set of feature data. A training sample set was established through image block extraction. Training data was annotated based on intact joint samples, and the connection status was classified into four categories to obtain the fifth set of feature data. A joint integrity score was calculated based on the fifth set of feature data. The edge strength score weight was determined using the entropy weight method. The weighted sum of the joint surface gap scores was performed. If the combined score fell below the preset threshold of 0.7, the joint was deemed to have an abnormality risk. Edge feature extraction is a key step in the inspection of shock-absorbing hammer joint components. The Sobel operator calculates the gradient by calculating the grayscale difference between the horizontal and vertical directions of the pixel. Taking a 3×3 operator as an example, the horizontal operator is [[-1,-2,-1], [0,0,0], [1,2,1]], and the vertical operator is [[-1,0,1], [-2,0,2], [-1,0,1]]. For images of normal joints, the edge gradient magnitude typically ranges from 150 to 200, while for loose or damaged joints, the gradient value drops below 80, and the gradient direction becomes disordered. For circular component detection, the Huff transform uses an accumulator matrix to detect circles. Assuming the center coordinates of the circle are (a, b) and the radius is r, the parametric equations of the circle are x = a + rcosθ, y = b + rsinθ. Typical circular components in shock-absorbing hammer connectors have a radius of 25 to 35 pixels. By setting the radius search range to [20, 40] and the accumulator threshold to 150, the connected components can be accurately detected. The distance between two circular components is between 5 and 8 pixels in a normal connection state, and a distance exceeding 10 pixels indicates an abnormal connection. The local binarization of the joints uses the Otsu algorithm, determining the threshold by maximizing the inter-class variance. For a 50×50 image of the connected region, the optimal threshold of 128 is calculated by calculating the grayscale histogram. Morphological processing using a 3×3 structuring element for dilation and erosion enhances the joint edges. Skeleton extraction uses a thinning algorithm to preserve the topological structure of the connection lines, and the skeleton lines at normal connections appear as regular straight lines or arcs.Edge strength is calculated using morphological gradients, equal to the difference between the dilated image and the eroded image. The mean edge strength for normal connections is around 45, with a standard deviation less than 8. However, for abnormal connections, the mean drops below 25 and the standard deviation exceeds 15. A distance transform is used to calculate the gap between connected surfaces. The maximum distance for normal connections does not exceed 3 pixels, and the average distance is around 1.5 pixels. The convolutional neural network uses a VGG architecture, consisting of five convolutional layers and three fully connected layers. The input is a 48×48 image patch, and the training samples are expanded to 2,000 through data augmentation. The network categorizes connection status into four categories: intact, slightly abnormal, moderately abnormal, and severely abnormal, achieving a training accuracy of 94%. In the comprehensive scoring, the weights for edge integrity, edge strength, and connection gap are 0.4, 0.35, and 0.25, respectively. The weighted scores for the three indicators for a shock-absorbing hammer connection were 0.82, 0.65, and 0.58, resulting in a weighted overall score of 0.68, which is below the threshold of 0.7 and indicates an abnormality risk.

[0039] S105. In response to the position deviation and abnormal working status of the shock-absorbing hammer, a transmission line inspection report is generated, and the satellite image and drone image of the abnormal position and the corresponding health assessment results are stored and sent to relevant personnel.

[0040] Abnormal shock-absorbing hammers are classified and labeled into four levels according to their position offset and health score data, and an abnormal position data set is obtained through geographic coordinate data. A natural language generation method is used to quantitatively describe the state parameters of the shock-absorbing hammers in the abnormal position data set, and an inspection anomaly assessment report is obtained through feature vector mapping. A spatiotemporal index table is established based on the inspection anomaly assessment report, and spatial data partitioning is used to identify abnormal positions in blocks, and hash algorithm encoding is used to obtain storage data blocks. A Bloom filter is used to check for duplicate verification on the stored data blocks. If the data is determined to be complete, the abnormal data is sorted according to the priority heap to obtain push data.

[0041] Specifically, based on the shock-absorbing hammer inspection results, abnormal shock-absorbing hammers were classified and labeled into four levels according to their position offset and health score. The abnormal locations were located using geographic coordinate data. Fusion images and assessment data were collected for each abnormal location, and the abnormal location data were spatially clustered based on latitude and longitude information to obtain the first set of abnormal data. For the first set of abnormal data, a natural language generation algorithm was used to construct a text description. The shock-absorbing hammer status parameters were quantitatively described through feature vector mapping. Abnormal assessment content was generated based on the inspection report template, and the assessment content was keyword-labeled to obtain the second set of inspection reports. A spatiotemporal index table was established based on the second set of inspection reports. Spatial data partitioning was used to identify abnormal locations in blocks. The data blocks were encoded and stored using a hash algorithm. The inspection data was distributedly backed up to obtain the third set of stored data. For the third set of stored data, a Bloom filter was used to check for duplicates and verify data integrity using a digital signature algorithm. Based on the verification results, the data was graded and labeled to obtain the fourth set of verified data. Based on the fourth set of verification data, a message queue is constructed. Four levels of abnormal data are sorted using a priority heap. A data push list is generated for different terminal devices, and the inspection data is sent in a tiered manner to produce the final push results. In processing shock-absorbing hammer abnormal data, a four-level classification is performed based on position offset and health score. A position offset less than 30 mm and a health score greater than 0.8 are labeled as minor abnormalities; a position offset between 30 and 50 mm or a health score between 0.6 and 0.8 are labeled as moderate abnormalities; a position offset between 50 and 80 mm or a health score between 0.4 and 0.6 are labeled as severe abnormalities; and a position offset exceeding 80 mm or a health score below 0.4 are labeled as critical abnormalities. Spatial clustering is performed based on the shock-absorbing hammer's geographic location using a density-based clustering algorithm. Clustering is performed based on latitude and longitude coordinates, with a cluster radius of 500 meters and a minimum number of points of three. For example, five anomalies were detected on a certain line, with coordinates of (116.3, 39.9), (116.4, 39.95), (116.35, 39.92), (117.2, 40.1), and (117.3, 40.15). After clustering, two anomaly areas were obtained. Natural language generation uses a template-based approach to construct descriptions through feature vector mapping. Taking the description of the shock absorber status as an example, the feature vector contains [position offset, health score, anomaly level, inspection time], which is mapped to the template "Inspection time [3] found that the shock absorber at position coordinates [longitude, latitude] had a level [2] anomaly, position offset [0] mm, and health score [1]". The eigenvector of a shock absorber is [55, 0.58, 3,"2024-03-15"], which generates the description text "A level 3 abnormality was detected in the shock absorber at the coordinates (116.35, 39.92) on 2024-03-15. The position offset is 55 mm, and the health score is 0.58."In the data storage phase, spatial indexing is used to accelerate queries. Anomaly locations are indexed using an R-tree structure, with leaf nodes storing the location boundaries and data pointers for the shock absorbers. Data segmentation utilizes a grid method, dividing the entire line area into 1 km x 1 km grids, each assigned a unique number. The grid numbers are mapped using the hash function H(x,y) = (x*10000 + y) % 1000 to obtain the storage location. Data integrity is verified using a Bloom filter and digital signature. The Bloom filter uses three hash functions and an array length of 10,000, enabling duplicate data detection with a false positive rate of 0.1%. The digital signature utilizes the RSA algorithm, generating a 128-bit signature for each piece of data. The message queue is implemented using a priority heap, with critical anomalies receiving a priority of 4, severe anomalies 3, moderate anomalies 2, and minor anomalies 1. Alarms are delivered to different terminals based on their level of severity, with critical anomalies delivered in real time and other levels delivered on a scheduled basis.

[0042] S106. During the subsequent transmission line inspection process, the anti-vibration hammer spectral feature library and image distortion correction method are continuously updated, and the accuracy of anti-vibration hammer position offset detection and working status assessment is continuously improved through incremental learning, forming a dynamically updated and optimized transmission line monitoring system.

[0043] A shock-absorbing hammer sample is obtained according to a sample database, and similarity matching is performed on the shock-absorbing hammer sample through spectral angle calculation. First feature data is obtained for samples whose similarity is lower than a preset threshold; based on the first feature data, the least squares method is used to optimize and calculate the distortion correction parameters, and the corrected image is evaluated by the structural similarity index to obtain second correction data; based on the second correction data, a verification sample set is established, and the detection accuracy is calculated by a five-fold cross-validation method. A detection curve is plotted for the true positive rate and the false positive rate to obtain third performance data; for the third performance data, features are quantitatively evaluated by information gain comparison, and key features are screened by a recursive feature elimination method to obtain an optimized feature library.

[0044] Specifically, new anti-vibration hammer samples were retrieved from the database based on the inspection cycle. Sample features were then similarly matched using spectral angle calculations. Feature extraction was performed on samples with matching similarities below a threshold of 0.8. The classifier parameters were incrementally updated using an online sequential extreme learning algorithm to obtain the first set of feature data. For this first set of feature data, the least squares method was used to optimize the distortion correction parameters. The quality of the corrected images was assessed using the structural similarity index. The convolutional neural network weights were then updated online based on the assessment scores to obtain the second set of corrected data. A validation sample set was constructed based on the second set of corrected data. The detection accuracy was calculated using a five-fold cross-validation method. A plot of the true positive rate against the false positive rate was plotted, and the detection performance was evaluated using the area under the curve to obtain the third set of performance data. Feature importance was calculated for the third set of performance data, and features were quantitatively evaluated using information gain comparisons. Dimensionality reduction was performed on the features based on the assessment scores to obtain the fourth set of feature data. A feature importance ranking table was established based on the fourth set of feature data. Recursive feature elimination was used to screen key features. A spectral feature library was reconstructed for the screened features to obtain the fifth set of optimized data. Adaptive threshold update is performed on the fifth set of optimized data, and the threshold parameters are optimized by dynamic programming algorithm. The detection accuracy is monitored online to obtain the final optimization result. In the dynamic optimization of the anti-vibration hammer detection model, the incremental learning process adopts the sequential limit learning algorithm. The novelty of the sample is judged by calculating the spectral angle of the new sample and the feature library sample. The spectral angle cos(θ)=(A·B) / (|A|·|B|), where A and B are the spectral feature vectors to be compared. When the spectral feature vector of a certain anti-vibration hammer sample is [0.6, 0.7, 0.8], the spectral angle of the closest sample in the feature library is 0.75, which is lower than the threshold of 0.8, indicating that the sample contains new feature information. In the distortion correction optimization, the least squares method is used to update the correction parameters. Assume that the distortion model is r'=r(1+k1 r 2 +k2r 4 ), where r is the distance from the pixel in the original image to the center of the image, r' is the distance after distortion correction, and k1 and k2 are the distortion coefficients. By minimizing the residual sum of squares Σ(r'-r) 2Update the parameters. The initial coefficients of a certain group of images are k1=0.05 and k2=-0.02. After optimization, they become k1=0.048 and k2=-0.018. The structural similarity index SSIM increased from 0.92 to 0.95, indicating that the correction quality has been improved. ROC curve analysis is used for performance evaluation. The true positive rate TPR and false positive rate FPR at different working points are calculated by adjusting the detection threshold. In the five-fold cross validation, each fold contains 200 samples, and the obtained TPRs are 0.92, 0.94, 0.91, 0.93, and 0.95, respectively, and the FPRs are 0.08, 0.07, 0.09, 0.06, and 0.08, respectively. The area under the curve AUC=0.93 is calculated by the trapezoidal integration method, verifying that the performance of the detector is stable. The feature importance is evaluated using the information gain ratio. For the 10 features of the shock-absorbing hammer, the calculated importance is

[0045] [0.85, 0.78, 0.72, 0.65, 0.58, 0.52, 0.45, 0.38, 0.32, 0.25]. Through recursive feature elimination, features with an importance exceeding 0.6 were selected to reconstruct the feature library, ultimately retaining five key features. During adaptive threshold updating, a dynamic programming algorithm was used to optimize the detection threshold. In one optimization process, the initial threshold was 0.7. After five rounds of iteration, the thresholds were adjusted to 0.68, 0.65, 0.66, 0.67, and 0.65, respectively. Detection accuracy increased from 0.91 to 0.94. Through online monitoring, a new round of optimization was triggered when the accuracy dropped by more than 3%.

[0046] S107. Regularly conduct video surveillance of satellite remote sensing images and drone inspection images of transmission lines, and perform batch processing and analysis of video surveillance data to identify shock-absorbing hammer targets in real time in the video, extract their motion trajectory and state change information, and integrate and verify them with static image analysis results. By analyzing the correlation patterns and evolution trends of shock-absorbing hammer failures, feedback is provided on the video surveillance identification content.

[0047] Receive inspection video data, perform motion target segmentation on adjacent frame images using an inter-frame difference algorithm, use a convolutional neural network to identify the target position of the shock-absorbing hammer, and extract target features based on the identification results to obtain a first set of dynamic data; for the first set of dynamic data, use a Kalman filter to construct a state prediction equation, use the optical flow method to calculate the displacement vector of the shock-absorbing hammer, and calculate the swing parameters based on the displacement vector to obtain a second set of tracking data; based on the second set of tracking data, use Fourier transform to decompose the motion signal, use multi-scale wavelet transform to reduce the noise of the state sequence, and extract spectral components based on the swing characteristics of the shock-absorbing hammer to obtain a third set of spectrum data; for the third set of spectrum data, use time series correlation analysis to extract characteristic rules, build a fault model through feature vectors, and predict the motion characteristics of the shock-absorbing hammer to obtain a fault prediction result.

[0048] Specifically, based on inspection video data, the inter-frame difference algorithm is used to segment moving targets between adjacent frames. A convolutional neural network is used to identify and locate the position of the shock absorber in each frame. Based on the identification results, three features, namely target size, position coordinates, and motion speed, are extracted. A time series feature sequence is established for the shock absorber target, generating the first set of dynamic data. For this first set of dynamic data, a Kalman filter is used to construct a target state prediction equation. The optical flow method is used to calculate the target displacement vector between adjacent frames. The shock absorber swing amplitude and frequency are calculated based on the motion vectors, and the state change parameters are recorded to generate the second set of tracking data. Based on the second set of tracking data, the motion periodic signal is decomposed using a Fourier transform. The state sequence is denoised using a multi-scale wavelet transform. The fundamental and harmonic components of the shock absorber swing characteristics are extracted, and the motion features are quantified to generate the third set of spectral data. Based on the third set of spectral data, temporal correlation analysis is used to extract motion feature patterns. A fault feature library is constructed using feature vectors, and cluster analysis of the shock absorber motion patterns is performed to generate the fourth set of regularity data. Based on the fourth set of regular data and static image analysis results, the association rule mining algorithm is used to extract feature association rules, and the reliability of the rules is verified by confidence calculation. The association rules are screened to obtain the fifth set of verification data. Based on the fifth set of verification data, a fault evolution model is constructed, and the fault development trend is predicted by the time series prediction algorithm. The recognition threshold is corrected based on the prediction result, and the recognition result is optimized by feedback to obtain the final analysis result. In the video monitoring analysis of the shock-absorbing hammer, the inter-frame difference is used to extract the moving target by calculating the pixel difference between adjacent frames. The threshold segmentation method is used, and the threshold T = 25 is set. When the pixel difference is greater than T, it is determined to be a moving area. For a video with 25 frames per second, the time interval between adjacent frames is 40 milliseconds. The displacement of the shock-absorbing hammer during this time is usually between 3 and 5 pixels. The noise effect is eliminated by using a 3×3 median filter. The Kalman filter is used for target tracking. The state vector contains position and velocity X = [x, y, vx, vy], and the state transfer matrix

[0049] F=[[1,0,dt,0],[0,1,0,dt],[0,0,1,0],[0,0,0,1]], where dt is the time interval. The process noise covariance Q and the measurement noise covariance R are determined experimentally.

[0050] Q = diag([0.1, 0.1, 0.2, 0.2]), R = diag([2, 2]). diag([0.1, 0.1, 0.2, 0.2]) indicates that Q is a diagonal matrix with diagonal elements of 0.1, 0.1, 0.2, and 0.2, respectively. diag([2, 2]) indicates that R is also a diagonal matrix with diagonal elements of 2. For a properly functioning shock-absorbing hammer, the error between the predicted and actual positions is within 2 pixels, and the velocity prediction error is less than 0.5 pixels per frame. In spectrum analysis, fast Fourier transform (FFT) is used to extract motion features. The 1024-point sampling sequence is transformed to obtain the fundamental frequency and harmonic components. The fundamental frequency of a normal shock-absorbing hammer is between 0.5 and 2 Hz, with the main harmonic components no more than 4 times the order and an amplitude ratio less than 0.3. The signal is decomposed into four layers using the db4 wavelet, and high-frequency noise is suppressed using a soft thresholding method. The reconstructed signal-to-noise ratio is improved by 8 to 12 decibels. Feature association analysis uses the Apriori algorithm to mine association rules. The minimum support and confidence levels are set to 0.2 and 0.8, respectively. For a shock-absorbing hammer, the rule "IF the vibration frequency increases AND the amplitude increases THEN the loosening risk increases" is extracted with a support of 0.25 and a confidence of 0.85. This rule is combined with the "bolt wear" feature obtained from static image analysis to form a complete fault pattern. Fault prediction uses a time series model, building an autoregressive model for the feature sequence. The model order is p = 4, and the parameters are estimated using least squares. For a fault evolution sequence, the health index gradually decreases from 0.95 to 0.75, with a predicted probability of 0.80 of decreasing to 0.65 within the next 72 hours. Based on the prediction results, the monitoring frequency of the shock-absorbing hammer is adjusted from 24 hours to 8 hours, and the recognition threshold is adjusted from 0.7 to 0.75. Through incremental training, the model's prediction accuracy improves from an initial 0.85 to 0.92.

[0051] The above content is merely an example and explanation of the structure of the present invention. Those skilled in the art may make various modifications or additions to the described specific embodiments or replace them in a similar manner. As long as they do not deviate from the structure of the invention or exceed the scope defined by the claims, they should all fall within the scope of protection of the present invention.

Claims

1. A video surveillance recognition method integrating satellite and UAV images, characterized in that: The method comprises: Obtain satellite and drone images, perform multi-scale fusion of the satellite and drone images using pyramid decomposition, and combine the high-frequency details and low-frequency contour information of the two images through weighted averaging to obtain transmission line images that meet the target resolution and have texture details. Analyze normal and offset shock-absorbing hammer image samples in historical data, extract spectral features reflecting position offset, build a shock-absorbing hammer spectral feature library based on the spectral features reflecting position offset, perform spectral matching on the fused image, calculate the spectral angle between the fused image and the samples in the feature library, and determine the precise position of the shock-absorbing hammer based on the spectral angle. If the minimum spectral angle is greater than the preset spectral angle threshold, it is determined that the shock-absorbing hammer has position offset. For images of the shock absorber position in transmission line images that meet the target resolution and have texture details, a deep learning-based image distortion correction model is used. By training on a large number of distorted image samples, the distorted area is adaptively corrected to obtain the corrected shock absorber image. The texture, shape, and color of the shock-absorbing hammer in the calibrated shock-absorbing hammer image are extracted and combined with the hammer's attribute information to construct a multi-dimensional shock-absorbing hammer health assessment index system. The support vector machine algorithm is used to classify the working status of the shock-absorbing hammer, assess the weight of the impact of the shock-absorbing hammer's working status on the health of the shock-absorbing hammer, and calculate the comprehensive health score of the shock-absorbing hammer. If the score is lower than the preset scoring threshold, the shock-absorbing hammer is judged to have an abnormal risk. In the event of position deviation or abnormal working status of the shock-absorbing hammer, a transmission line inspection report is generated, and satellite and drone images of the abnormal location and the corresponding health assessment results are stored and sent to relevant personnel.

2. The method according to claim 1, characterized in that The method includes acquiring satellite images and drone images, performing multi-scale fusion of the satellite images and the drone images using a pyramid decomposition method, combining high-frequency details and low-frequency contour information of the two images by weighted averaging, and obtaining a transmission line image having a target resolution and texture details. The method includes: Acquire satellite image data of a power transmission line area, perform geometric correction on the satellite image data using a nearest neighbor resampling method, and use a median filter to eliminate image noise to obtain first image data; According to the geographic spatial reference information in the first image data, the image collected by the drone is geometrically corrected using a bilinear interpolation method, and noise reduction processing is performed using a Gaussian filter to obtain second image data; Performing pyramid decomposition on the first image data and the second image data, extracting image contour edge features using a Laplacian operator, and setting a wavelength threshold according to grayscale value differences to obtain high-frequency components and low-frequency components; For high-frequency components, local area variance is used as the weight coefficient for fusion, and for low-frequency components, local energy is used as the weight coefficient for fusion. The fused data is reconstructed by inverse pyramid transform to obtain the target image.

3. The method according to claim 1, characterized in that The method includes analyzing normal and offset shock-absorbing hammer image samples in historical data, extracting spectral features reflecting position offset, constructing a shock-absorbing hammer spectral feature library based on the spectral features reflecting position offset, performing spectral matching on the fused image, calculating the spectral angle between the fused image and the samples in the feature library, determining the precise position of the shock-absorbing hammer based on the spectral angle, and determining that the shock-absorbing hammer has position offset if the minimum spectral angle is greater than a preset spectral angle threshold. receiving an anti-vibration hammer image, performing feature enhancement on the image through RGB channel space conversion, processing the image with a Gaussian filter to obtain a preprocessed image, and extracting reflectivity, texture, and edge mutation spectral features from the preprocessed image to obtain feature data; Standardizing the characteristic data to eliminate dimensional effects, calculating a characteristic covariance matrix through principal component transformation, and performing eigenvalue decomposition on the covariance matrix to obtain a position characteristic matrix; A Gaussian kernel function is used to perform mapping transformation on the position feature matrix, a support vector machine algorithm is used to construct a feature space of the normal position and offset position of the shock-absorbing hammer, and a spectral feature library of the shock-absorbing hammer is established; The spectral angle is calculated based on the feature space, and the spatial position coordinates of the shock-absorbing hammer are fitted using the least squares method to obtain a position offset. If the spectral angle is greater than a preset spectral angle threshold and the position offset is greater than a preset distance threshold, it is determined that the shock-absorbing hammer has a position offset.

4. The method according to claim 1, wherein The method includes: using a deep learning-based image distortion correction model to adaptively correct the distorted area of ​​the transmission line image at the location of the shock absorber in the transmission line image having target resolution and texture details by training a large number of distorted image samples to obtain a corrected shock absorber image. Obtaining a distorted image sample of the anti-vibration hammer, performing noise reduction processing on the distorted image sample through a Gaussian filter, and obtaining a first set of feature data using a radial distortion and a tangential distortion mapping function; Extracting image features using a convolutional neural network based on the first set of feature data, and iteratively updating network parameters using a back-propagation algorithm to obtain training parameters; Based on the training parameters, the feature map is upsampled by bilinear interpolation, and a corrected image is obtained by using an adaptive compensation function; Edge features are extracted from the corrected image, straight line boundaries are detected through Hough transform, and the least squares method is used to fit the contour curve to obtain the center coordinates of the shock-absorbing hammer. A geometric transformation matrix is ​​established based on the center coordinates of the shock-absorbing hammer to correct the distorted area.

5. The method according to claim 1, wherein The texture, shape, and color of the shock absorber are extracted from the corrected shock absorber image and combined with the attribute information of the shock absorber to construct a multi-dimensional shock absorber health evaluation index system. The working status of the shock absorber is classified using a support vector machine algorithm, the weight of the impact of the working status of the shock absorber on the health of the shock absorber is evaluated, and a comprehensive health score of the shock absorber is calculated. If the score is lower than a preset score threshold, it is determined that the shock absorber has an abnormal risk, including: Obtain the gray-level co-occurrence matrix and third-order color moment of the shock-absorbing hammer image, and obtain the first eigenvector by calculating the contrast, entropy, moment of inertia, and surface color distribution; Calculating attribute thresholds of corrosion degree, crack degree, and integrity degree using the maximum inter-class variance method based on the first eigenvector, and obtaining a second eigenvector using a region growing algorithm; Acquire historical data on the use time, quality parameters, and vibration frequency of the shock-absorbing hammer for the second eigenvector, and select eigenvectors whose cumulative contribution rate exceeds a preset threshold through principal component analysis to obtain a third eigenvector; A support vector machine is used to construct a classifier based on the third eigenvector, and a kernel function parameter is optimized through grid search to obtain a health score of the shock-absorbing hammer. If the health score is lower than a preset score threshold, it is determined that the shock-absorbing hammer has an abnormal risk.

6. The method according to claim 1, characterized in that The method extracts the texture, shape, and color of the shock-absorbing hammer from the corrected shock-absorbing hammer image and combines this with attribute information of the shock-absorbing hammer to construct a multi-dimensional shock-absorbing hammer health assessment index system. The system classifies the working status of the shock-absorbing hammer using a support vector machine algorithm, assesses the weight of the impact of the working status of the shock-absorbing hammer on the health of the shock-absorbing hammer, and calculates a comprehensive health score of the shock-absorbing hammer. If the score is lower than a preset score threshold, the shock-absorbing hammer is determined to have an abnormal risk. The method includes: applying an edge detection method to the corrected shock-absorbing hammer image to highlight the boundaries and contours of the shock-absorbing hammer components, using a shape recognition method to identify shapes corresponding to connected components, analyzing pixel intensity gradients near the shapes corresponding to the connected components, accurately locating connections between components, and assessing a comprehensive health score of the shock-absorbing hammer connection. If the score is lower than a preset score threshold, the shock-absorbing hammer connection is determined to have an abnormal risk.

7. The method according to claim 1, characterized in that The above mentioned system generates a transmission line inspection report for situations where the anti-vibration hammer is misaligned or in an abnormal working state, and stores and sends satellite and drone images of the abnormal position and corresponding health assessment results to relevant personnel, including: Abnormal shock absorbers are classified and labeled into four categories based on their position offset and health score data, and the abnormal position dataset is obtained through geographic coordinate data. A natural language generation method is used to quantitatively describe the state parameters of the shock-absorbing hammer in the abnormal position data set, and an inspection abnormality assessment report is obtained through feature vector mapping; Establish a spatiotemporal index table based on the inspection anomaly assessment report, use spatial data partitioning to identify the anomaly location in blocks, and obtain storage data blocks through hash algorithm encoding; A Bloom filter is used to check for duplicates in the stored data blocks. If the data is determined to be complete, the abnormal data is sorted according to a priority heap to obtain pushed data.

8. The method according to claim 1, characterized in that The method further includes: continuously updating the anti-vibration hammer spectral feature library and image distortion correction method during subsequent transmission line inspections, continuously improving the accuracy of anti-vibration hammer position offset detection and working status assessment through incremental learning, and forming a dynamically updated and optimized transmission line monitoring system; Regular video monitoring of satellite remote sensing images and drone inspection images of transmission lines is carried out, and the video monitoring data is batch processed and analyzed to identify shock-absorbing hammer targets in real time in the video, extract their motion trajectory and state change information, and integrate and verify them with static image analysis results. By analyzing the correlation patterns and evolution trends of shock-absorbing hammer failures, the video monitoring identification content is fed back.

9. The method according to claim 8, characterized in that During subsequent transmission line inspections, the shock-absorbing hammer spectral feature library and image distortion correction method are continuously updated, and the accuracy of shock-absorbing hammer position offset detection and working status assessment is continuously improved through incremental learning, forming a dynamically updated and optimized transmission line monitoring system, including: Acquire shock-absorbing hammer samples from a sample database, perform similarity matching on the shock-absorbing hammer samples by spectral angle calculation, and obtain first feature data for samples whose similarity is lower than a preset threshold; Based on the first feature data, the least square method is used to optimize and calculate the distortion correction parameters, and the corrected image is evaluated by the structural similarity index to obtain the second correction data; Establishing a validation sample set based on the second calibration data, calculating the detection accuracy using a five-fold cross-validation method, and plotting a detection curve based on the true positive rate and the false positive rate to obtain third performance data; For the third performance data, features are quantitatively evaluated through information gain comparison, and recursive feature elimination method is used to screen key features to obtain an optimized feature library.

10. The method according to claim 8, characterized in that The system regularly monitors satellite remote sensing images and drone inspection images of power transmission lines, processes and analyzes the video monitoring data in batches, identifies the shock absorber targets in the video in real time, extracts their motion trajectory and state change information, and integrates and verifies the results with static image analysis. By analyzing the correlation patterns and evolution trends of shock absorber faults, the system provides feedback on the video monitoring identification content, including: Receive inspection video data, perform moving target segmentation on adjacent frame images using an inter-frame difference algorithm, use a convolutional neural network to identify the target position of the shock-absorbing hammer, and extract target features based on the recognition results to obtain the first set of dynamic data; For the first set of dynamic data, a Kalman filter is used to construct a state prediction equation, an optical flow method is used to calculate the displacement vector of the shock-absorbing hammer, and a swing parameter is calculated based on the displacement vector to obtain a second set of tracking data; Based on the second set of tracking data, the motion signal is decomposed using Fourier transform, the state sequence is denoised using multi-scale wavelet transform, and the spectral components are extracted based on the swing characteristics of the shock-absorbing hammer to obtain a third set of spectral data; For the third set of spectrum data, time series correlation analysis is used to extract characteristic patterns, a fault model is constructed through characteristic vectors, and the motion characteristics of the anti-vibration hammer are predicted to obtain a fault prediction result.

Images