A power transmission and distribution construction monitoring method and system based on air-ground communication
By using air-to-ground communication technology and image processing methods, construction process nodes can be identified in real time and standardized image sequences can be generated, which solves the problems of low efficiency and noise impact in power transmission and distribution construction monitoring and enables accurate construction safety judgment.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHANDONG JIAYU CONSTR ENG CO LTD
- Filing Date
- 2026-02-27
- Publication Date
- 2026-05-29
AI Technical Summary
Existing power transmission and distribution construction monitoring technologies rely on manual inspections and single image acquisition, which are inefficient, have blind spots, and are affected by image noise. They are difficult to achieve real-time monitoring of the entire process and the entire cycle, resulting in large deviations in the identification of construction processes and inaccurate extraction of monitoring target features.
A construction monitoring method based on air-to-ground communication is adopted. By deploying image acquisition equipment in a coordinated air-to-ground manner, construction process nodes are identified in real time, static ground feature anchor points that are fixed within the frame are extracted, rigid alignment is performed to generate standardized image sequences, and gray-level gradient and edge texture feature vectors are compared to achieve accurate judgment.
It improves the accuracy of construction process node identification, effectively avoids image noise and spatial misalignment problems, realizes accurate and efficient judgment of construction safety operations, and enhances the intelligence level and reliability of power transmission and distribution construction monitoring.
Smart Images
Figure CN122116279A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of power construction technology, and in particular relates to a method and system for monitoring power transmission and distribution construction based on air-to-ground communication. Background Technology
[0002] Power transmission and distribution engineering construction scenarios are complex, often involving high-altitude and field operations. Construction procedures are closely linked, and stringent requirements are placed on operational standardization and quality control. Real-time and accurate construction monitoring is a core element in ensuring project safety and quality. Currently, power transmission and distribution construction monitoring largely relies on manual inspections and single image acquisition monitoring. Manual inspections are inefficient, have blind spots, and are difficult to implement real-time monitoring of the entire process and lifecycle. Traditional image monitoring technologies are easily affected by factors such as the construction environment and equipment movement, resulting in image noise, spatial misalignment, and other problems, leading to significant errors in identifying construction procedures and inaccurate extraction of monitored target features. Summary of the Invention
[0003] In view of the technical problems existing in the above-mentioned background art, the present invention proposes a method and system for monitoring power transmission and distribution construction based on air-to-ground communication.
[0004] To achieve the above objectives, the technical solution adopted by the present invention includes the following steps:
[0005] S1. Acquire the initial image sequence of the construction site through image acquisition equipment deployed in a coordinated air-ground manner, and identify the construction process nodes in real time based on the initial image sequence;
[0006] S2. For each frame image in the initial image sequence, extract the feature anchor points corresponding to the static ground features that remain unchanged within the frame;
[0007] S3. Perform rigid alignment based on feature anchor point matching to generate a standardized image sequence with spatial alignment throughout the entire construction cycle;
[0008] S4. Based on standardized image sequences, delineate dynamic construction monitoring target areas that match the current process node, extract grayscale gradient feature vectors and edge texture feature vectors within the target area, and synchronously transmit the extracted feature vectors to the control terminal via an air-to-ground communication link.
[0009] S5. Based on the received feature vector, the control terminal retrieves the pre-stored standardized construction feature benchmarks for the corresponding process node, compares the current construction feature vector with the standardized construction feature benchmarks, and completes an accurate judgment on safe operation.
[0010] Preferably, the implementation of real-time identification of construction process nodes based on the initial image sequence in step S1 includes:
[0011] S11. For each frame of the initial image sequence, extract the personnel operation vector and the device status vector from each frame.
[0012] S12. Pre-construct a two-vector prior logic graph of power transmission and distribution construction procedures. The two-vector prior logic graph includes at least the personnel operation reference vector and equipment status reference vector corresponding to each standard construction procedure, as well as the sequential dependence of each procedure, the shortest / longest duration constraint of a single procedure, and the precondition constraint for starting the procedure.
[0013] S13. Perform similarity matching between the personnel operation vector and the personnel operation reference vector of each process in the dual-vector prior logic graph, and output the first preliminary process result and the first confidence level; at the same time, perform similarity matching between the equipment state vector and the equipment state reference vector of the corresponding process in the dual-vector prior logic graph, and output the second preliminary process result and the second confidence level.
[0014] S14. By using an attention mechanism to weightedly fuse the first and second confidence levels, the process with the highest overall matching degree is output as the final process node.
[0015] Preferably, before step S2, each frame in the initial image sequence needs to undergo image denoising, including:
[0016] S211. For each frame of image, the local pixel range and the minimum eigenvalue of the first-order structure tensor are weighted and fused to calculate the pixel-level noise variance.
[0017] S212. Construct a spatial structure tensor and a frequency domain amplitude-phase decoupling tensor respectively. The spatial structure tensor includes a first-order structure tensor and a second-order structure tensor.
[0018] S213. Perform a two-dimensional Fourier transform on the image to decouple the amplitude and phase to obtain the phase tensor and the amplitude tensor. Keep the phase tensor unchanged, use the noise variance to correct the amplitude tensor, and combine the corrected amplitude tensor with the phase tensor to obtain the frequency domain denoised image through inverse Fourier transform.
[0019] S214. Using the eigenvalue vectors of the first-order structure tensor and the second-order structure tensor as the basis, calculate the weights of the non-local image blocks and perform weighted filtering to obtain the spatial domain denoised image.
[0020] S215. Based on the eigenvalues of the first-order structure tensor, an adaptive fusion weight is constructed. The weights are dynamically allocated according to the local structural features of the construction image. The frequency domain and spatial domain denoising results are fused to obtain the denoised image.
[0021] Preferably, step S214, which uses the eigenvectors of the first-order structure tensor and the second-order structure tensor as the similarity basis to calculate the weights of non-local image patches and performs weighted filtering to obtain a spatially denoised image, includes:
[0022] S2141. First, calculate the tensor similarity between any two image patches within the nonlocal search window;
[0023] S2142. Normalize the calculated similarity to obtain the weighting coefficients for each image block, ensuring that the sum of the weights is 1.
[0024] S2143. Based on the weighting coefficients, perform weighted fusion on all image blocks within the non-local search window to obtain a spatially structured denoised image.
[0025] Preferably, step S2, which extracts feature anchor points corresponding to static features that remain unchanged within each frame of the initial image sequence, includes the following steps:
[0026] S221. For each frame of the initial image sequence, first identify the preset static ground object categories in the power transmission and distribution construction scene through a lightweight semantic segmentation network, and generate the corresponding semantic mask.
[0027] S222. Based on the geometric prior rules of static features, the semantic mask is filtered a second time to remove the falsely detected dynamic interference areas and obtain the static feature candidate areas.
[0028] S223. Perform edge detection on each static ground feature candidate region and then extract the geometric vertex set of the candidate region through a polygon fitting algorithm;
[0029] S224. Perform corner detection algorithm on each static ground feature candidate area to extract the corner set;
[0030] S225. Perform spatial distance clustering on the geometric vertex set and corner set, and use the cluster centers where the spatial distance between the two types of feature points is less than a preset threshold as feature anchor points.
[0031] Preferably, step S3, which performs rigid alignment based on feature anchor point matching to generate a standardized image sequence with spatial alignment throughout the entire construction cycle, includes:
[0032] S31. Using the first frame after denoising as the reference frame, establish an initial set of anchor points corresponding to the reference frame and the current frame to be aligned based on feature anchor points.
[0033] S32. Perform triangulation on the anchor point set in the reference frame to generate a topological triangle network. For each topological triangle, calculate its topological invariant vector in the reference frame. For the corresponding triangle in the current frame to be aligned, calculate its topological invariant vector and calculate the Euclidean distance between them. If the Euclidean distance is greater than the set topological error threshold, it is determined that there is an abnormal match of the anchor points in the triangle. The corresponding anchor point pair is removed to generate the final reliable matching set.
[0034] S33. According to the category of the feature anchor points, the reliable matching set is divided into L subsets. For each subset, the random sampling consensus algorithm combined with the least squares method is used to estimate the local homography matrix corresponding to the land cover of that category, and the mean value of the anchor point reprojection error of that subset is calculated.
[0035] S34. Calculate the adaptive weights of each local homography matrix based on the mean reprojection error, and fuse all local homography matrices according to the adaptive weights to obtain the global rigid alignment homography matrix.
[0036] S35. Perform a complete perspective transformation resampling using a global rigid alignment homography matrix to obtain the pixel coordinates of the aligned image and complete the standardized image sequence.
[0037] Preferably, step S4, which involves extracting the grayscale gradient feature vector and edge texture feature vector within the target region, includes:
[0038] S41. Gray-level gradient feature vector extraction: For the defined target region ROI, a local pixel coordinate system with the associated feature anchor point as the origin is constructed. Multi-directional gray-level gradient calculation and normalization are performed on each pixel in the ROI to generate a global gray-level gradient feature vector.
[0039] S42. For the defined target region ROI, calculate multi-scale texture features with edge pixels as the core, and generate an edge texture feature vector that matches the dimension of the gray-level gradient feature vector.
[0040] S43. Based on the temporal continuity of the standardized image sequence, weighted moving average fusion is performed on the gray-level gradient feature vector and edge texture feature vector of the same ROI in three consecutive frames to remove feature mutations caused by noise and obtain the final output feature vector.
[0041] Preferably, the comparison between the current construction feature vector and the standardized construction feature benchmark in step S5 is obtained by weighted cosine similarity calculation.
[0042] As a preferred embodiment, a power transmission and distribution construction monitoring system based on air-to-ground communication includes: an image acquisition and recognition module, a feature anchor point extraction module, a standardized image sequence module, a feature extraction and transmission module, and an intelligent judgment module.
[0043] The image acquisition and recognition module acquires the initial image sequence of the construction site through image acquisition equipment deployed in a coordinated air-ground manner, and identifies the construction process nodes in real time based on the initial image sequence;
[0044] The feature anchor point extraction module extracts the feature anchor points corresponding to the static ground features that remain unchanged within each frame of the initial image sequence.
[0045] The standardized image sequence module performs rigid alignment based on feature anchor point matching to generate a standardized image sequence with spatial alignment throughout the entire construction cycle.
[0046] The feature extraction and transmission module is based on a standardized image sequence, delineates a dynamic construction monitoring target area that matches the current process node, extracts gray-level gradient feature vectors and edge texture feature vectors within the target area, and synchronously transmits the extracted feature vectors to the control terminal through an air-to-ground communication link.
[0047] The intelligent judgment module is a control terminal that, based on the received feature vector, retrieves the pre-stored standardized construction feature benchmarks for the corresponding process node, compares the current construction feature vector with the standardized construction feature benchmarks, and completes an accurate judgment on safe operation.
[0048] Compared with existing technologies, the advantages and positive effects of this invention are as follows: Relying on a dual-vector prior logic graph and attention mechanism, this invention significantly improves the accuracy of construction process node identification and effectively avoids the misjudgment problem of single feature matching. Simultaneously, through frequency-domain and spatial-domain joint adaptive denoising and rigid feature anchor point alignment technology, it solves the industry pain points of image noise and spatial misalignment in construction scenarios, generating standardized image sequences. Combined with process-adaptive feature extraction and weighted similarity comparison, it achieves accurate and efficient judgment of construction safety operations, comprehensively improving the intelligence level and control reliability of power transmission and distribution construction monitoring. Attached Figure Description
[0049] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0050] Figure 1 This is a flowchart of a power transmission and distribution construction monitoring method and system based on air-to-ground communication. Detailed Implementation
[0051] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described below in conjunction with the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.
[0052] Numerous specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways than those described herein, and therefore the invention is not limited to the specific embodiments disclosed in the following specification.
[0053] This invention discloses a method and system for monitoring power transmission and distribution construction based on air-to-ground communication, as described in the embodiments below. Figure 1 .
[0054] First, an initial image sequence of the construction site is acquired using image acquisition equipment deployed in a coordinated air-ground manner. Based on this initial image sequence, construction process nodes are identified in real time. The image acquisition equipment includes devices such as drones. The implementation of real-time identification of construction process nodes based on the initial image sequence includes: for each frame of the initial image sequence, extracting personnel operation vectors and equipment state vectors; pre-constructing a dual-vector prior logic graph for power transmission and distribution construction processes, which at least includes personnel operation reference vectors, equipment state reference vectors corresponding to each standard construction process, as well as the sequential dependencies between processes, the shortest / longest duration constraints for a single process, and the precondition constraints for process initiation; performing similarity matching between the personnel operation vectors and the personnel operation reference vectors of each process in the dual-vector prior logic graph, outputting a first preliminary process result and a first confidence level; simultaneously performing similarity matching between the equipment state vectors and the corresponding equipment state reference vectors of the processes in the dual-vector prior logic graph, outputting a second preliminary process result and a second confidence level; and using an attention mechanism to weightedly fuse the first and second confidence levels, outputting the process with the highest overall matching degree as the final process node. Specifically, the first step is to perform similarity matching between the personnel operation vector and the baseline vector, outputting the first preliminary process result and the first confidence score. This step uses the cosine similarity algorithm, which is more robust to action trend recognition, as the core matching algorithm. This algorithm can avoid the interference of individual differences in the range of operation amplitude of construction workers on the matching results, focusing only on the core feature trend matching of operation behavior. Before matching, the real-time extracted personnel operation vector and the personnel operation baseline vector of each standard process in the dual-vector prior logic graph are first normalized using a min-max method to eliminate the dimensional differences of parameters such as action amplitude and spatial coordinates. Then, the normalized real-time personnel operation vector is used to calculate the cosine similarity value with each of the personnel operation baseline vectors of all standard processes in the graph. The similarity value ranges from 0 to 1, with a value closer to 1 indicating a higher degree of matching of the operation behavior. After completing the full calculation, processes with similarity values higher than the preset personnel matching threshold are first selected as candidate processes on the personnel side. The process with the highest similarity is the first preliminary process result. If the matching result does not meet the pre-start conditions of the process, its confidence value is reduced according to preset rules, and the corrected similarity value is the first confidence value. Simultaneously, the similarity matching between the equipment state vector and the reference vector is completed, and the second preliminary process result and the second confidence value are output. Consistent with the personnel vector matching logic, the real-time equipment state vector and the equipment state reference vector of each process in the graph are first normalized, and then the cosine similarity value is calculated one by one. At the same time, strong constraint verification of equipment category is superimposed. If the real-time identified equipment type does not match the reference equipment type of the corresponding process, the similarity value of the process is directly reduced in weight.Processes with similarity scores higher than a preset equipment matching threshold are selected as candidate processes on the equipment side. The process with the highest similarity score is the second preliminary process result. The similarity value after being corrected by process duration constraints is the second confidence score. Finally, a weighted fusion of the two confidence scores is performed through an attention mechanism to determine the final process node. This step adopts an adaptive attention weight allocation mechanism instead of a fixed weight fusion mode. It can automatically adjust the weight ratio of the two confidence scores according to the process type. For example, for personnel-dominated processes such as insulator installation, the personnel confidence score weight is automatically increased; for equipment-dominated processes such as foundation pouring, the equipment confidence score weight is automatically increased. Through the attention weight generation network, the personnel confidence score weight and the equipment confidence score weight are output. The comprehensive matching degree of all candidate processes is calculated by weighted summation. After final verification based on the temporal dependency relationship of the graph, the process with the highest comprehensive matching degree is selected as the final process node of the current construction.
[0055] In addition, image denoising is required for each frame in the initial image sequence before feature anchor point extraction. Specifically, for each frame... The pixel-level noise variance is calculated by weighting and fusing the local pixel range with the minimum eigenvalue of the first-order structure tensor. The calculation method is as follows: ,in, For The local pixel difference centered on the pixel This is the weighting coefficient, with a value ranging from 0 to 1. For a first-order structure tensor, It is the smallest eigenvalue of the first-order structure tensor.
[0056] A spatial structure tensor and a frequency domain amplitude-phase decoupling tensor are constructed respectively. The spatial structure tensor includes a first-order gradient covariance tensor and a second-order Hessian covariance tensor, and their calculation formulas are as follows: , ,in, The image gradient vector, For the outer product of vectors, It is a Hessian matrix.
[0057] For each frame of image Each pixel position Performing a two-dimensional Fourier transform to decouple the amplitude and phase yields the result. ,in, For amplitude tensors, For the phase tensor; noise suppression is applied only to the amplitude tensor, preserving the phase tensor, and the corrected amplitude tensor. ,in, Given the magnitude of the frequency signal, the corrected amplitude tensor is combined with the phase tensor and transformed using the inverse Fourier transform operator to obtain the denoised image in the frequency domain. .
[0058] by Using the feature vectors as a similarity basis, the weights of non-local image patches are calculated and weighted filtering is applied to obtain a spatially denoised image. ,in, Let be the weighting coefficient of image patch q over image patch p. Let q be the pixel grayscale value of image patch q. Specifically, first, calculate the tensor similarity between any two image patches within the non-local search window. The higher the similarity, the closer the structures of the two image patches are, and the more similar their noise distributions are. The similarity metric formula is: ,in, The tensor similarity between image patch p and image patch q is defined by a value in the range (0,1). The closer the value is to 1, the more similar the structures and noise distributions of the two image patches are. p and q are any two image patches within the nonlocal search window W. The search window W is selected as 15×15 or 21×21, and the image patch size is selected as 3×3 or 5×5. They are first-order structure tensors Second-order structure tensor eigenvectors; These are the L2 norms of the first-order and second-order structure tensor eigenvectors of the two image patches, respectively, used to measure the structural differences between the two image patches.
[0059] Secondly, the calculated similarity is normalized to obtain the weighting coefficients for each image block, ensuring that the sum of the weights is 1 to avoid image grayscale distortion during weighted fusion. The weight normalization formula is as follows: .
[0060] Finally, based on the weighting coefficients, all image blocks within the non-local search window are weighted and fused to obtain a spatial structure denoised image. This image can effectively suppress spatial noise (such as blur noise caused by dust and motion noise caused by construction workers) while preserving the details of the construction structure.
[0061] An adaptive fusion weight is constructed based on the eigenvalues of the first-order structure tensor. The weights are dynamically allocated according to the local structural features of the construction image. The denoising results in the frequency domain and spatial domain are fused to obtain the final denoised frame image. .in, The adaptive fusion weights are determined by the ratio of the largest to the smallest eigenvalue of the first-order structure tensor, and are calculated using the following formula: .in, For first-order structure tensor The larger the eigenvalue, the more distinct the edge of the local region. In this case, it is necessary to focus on frequency domain phase fidelity denoising to avoid edge blurring; conversely, a smaller eigenvalue indicates a more distinct edge. When the value is small, it indicates that the area is a smooth region (such as the tower plane or the ground). In this case, it is necessary to focus on denoising the airspace structure and suppressing noise in the smooth region.
[0062] Then, for each frame in the initial image sequence, feature anchor points corresponding to static features that remain unchanged within the frame are extracted. Specifically, for each frame in the initial image sequence, a lightweight semantic segmentation network is first used to identify the preset static feature categories in the power transmission and distribution construction scenario, generating corresponding semantic masks. Based on the geometric prior rules of static features, the semantic masks are further filtered to remove falsely detected dynamic interference regions, resulting in candidate static feature regions. Edge detection is performed on each candidate static feature region, and then a polygon fitting algorithm is used to extract the geometric vertex set of the candidate region. A corner detection algorithm is performed on each candidate static feature region to extract the corner set. Spatial distance clustering is performed on the geometric vertex set and the corner set, and the cluster centers where the spatial distance between the two types of feature points is less than a preset threshold are used as feature anchor points.
[0063] After determining the feature anchor points, rigid alignment based on feature anchor point matching is performed to generate a standardized image sequence with spatial alignment throughout the entire construction cycle. Specifically, the first denoised image is used as the reference frame. Establish a reference frame and the current frame to be aligned based on feature anchor points. initial set of anchor points ,in Reference frames and the current frame to be aligned The anchor points, where K is the number of matched anchor points. anchor set in Perform Delaunay triangulation to generate a topological triangular network. , where M is the number of triangles. For each topological triangle (vertices are) ), calculate its position in the reference frame topological invariant vectors in ,in, As the ratio of side lengths is an invariant, The angles of the three interior angles of a triangle are invariant; for The corresponding triangle Calculate its topological invariant vector and calculate with If the Euclidean distance is greater than a set topological error threshold, then the anchor points within the triangle are considered to have abnormal matching, the corresponding anchor point pairs are removed, and a final reliable matching set is generated. .
[0064] Reliable matching sets are categorized according to the category of feature anchors. The data is divided into L subsets. For each subset, the Random Sample Consensus (RANSAC) algorithm combined with the least squares method is used to estimate the local homography matrix corresponding to that land cover type. And calculate the mean anchor point reprojection error of this subset. The adaptive weights of each local homography matrix are calculated based on the mean of the reprojection error. ,in, Let be the mean reprojection error of the l-th subset. All local homography matrices are fused using adaptive weights to obtain a global rigid alignment homography matrix. A complete perspective transformation resampling is then performed using the global rigid alignment homography matrix to obtain the pixel coordinates of the aligned image.
[0065] Next, based on the standardized image sequence, a dynamic construction monitoring target area matching the current process node is delineated, and gray-level gradient feature vectors and edge texture feature vectors within the target area are extracted. The extracted feature vectors are then synchronously transmitted to the control terminal via an air-to-ground communication link.
[0066] Specifically, firstly, based on the identified current process node, the pre-stored construction area prior rules bound to that process node are retrieved. Using the extracted static feature anchor points as fixed spatial references, a dynamic construction monitoring target area (ROI) matching the current process node is delineated in the current frame of the generated standardized image sequence. The construction area prior rules include the core range of the corresponding process's construction actions, the relative spatial position constraints with the feature anchor points, and the minimum bounding rectangle size constraints of the area. The delineated ROI has a completely fixed relative spatial position with the static feature anchor points in the standardized image sequence throughout the entire construction cycle.
[0067] Gray-level gradient feature vector extraction is performed. For the defined target region (ROI), a local pixel coordinate system is constructed with the associated feature anchor point as the origin. Multi-directional gray-level gradient calculation and illumination robust normalization are performed on each pixel within the ROI to generate a global gray-level gradient feature vector. Specifically, for each pixel within the ROI, the Sobel operator in four directions is used to calculate the gray-level gradient magnitude in four directions (0°, 45°, 90°, and 135°) to adapt to the structural characteristics of power transmission and distribution construction, thus constructing a single-pixel gradient vector. Using the average gradient magnitude of the region where the static feature anchor point associated with the ROI is located as a benchmark, the gradient vector of a single pixel is normalized to eliminate the interference of outdoor lighting changes on the features. The calculation method is as follows: ,in, The average gradient magnitude of the associated feature anchor region. To prevent division by zero and minima, the normalized single-pixel gradient vectors of all pixels within the ROI are flattened and concatenated sequentially according to the row and column order of the local pixel coordinate system to generate a global grayscale gradient feature vector with fixed dimensions.
[0068] Next, for the defined target region ROI, multi-scale texture features are calculated with edge pixels as the core, generating edge texture feature vectors that match the dimension of the grayscale gradient feature vector. Specifically, this is achieved by processing each pixel within the ROI. A multi-scale sliding window is constructed with it as the center, and the gray-level co-occurrence matrix features of each window are calculated, including contrast, entropy, correlation and energy. A single-pixel texture feature vector is constructed, and the constrained single-pixel edge texture feature vectors of all pixels in the ROI are flattened and stitched together in a completely consistent local pixel coordinate system row and column order to generate a global edge texture feature vector that matches the dimension of the gray-level gradient feature vector.
[0069] Based on the temporal continuity of standardized image sequences, weighted moving average fusion is performed on the gray-level gradient feature vectors and edge texture feature vectors of the same ROI in three consecutive frames to remove feature abrupt changes caused by noise, and the final output feature vector is obtained. The extracted feature vector is then synchronously transmitted to the control terminal through an air-to-ground communication link.
[0070] Finally, based on the received feature vector, the control unit retrieves the pre-stored standardized construction feature benchmarks for the corresponding process node, and compares the current construction feature vector with the standardized construction feature benchmarks to accurately determine the safety of the operation. Specifically, comparing the current construction feature vector with the standardized construction feature benchmarks involves performing a weighted similarity comparison calculation of two-dimensional features. A weighted cosine similarity algorithm is used as the core comparison algorithm. First, the cosine similarity between the real-time grayscale gradient feature vector and the standard grayscale gradient benchmark vector, and the cosine similarity between the real-time edge texture feature vector and the standard edge texture benchmark vector are calculated separately. Then, based on the construction characteristics of the current process, an adaptive weighting mechanism is used to assign weights to the two sets of similarities, ultimately calculating a comprehensive similarity value. The comprehensive similarity value is compared with a pre-stored quality qualification threshold. If the comprehensive similarity value is lower than the qualification threshold, it indicates that there is a potential risk in the operation, thus issuing an early warning and accurately determining the safety of power transmission and distribution construction operations.
[0071] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments for application in other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.
Claims
1. A method for monitoring power transmission and distribution construction based on air-to-ground communication, characterized in that, Includes the following steps: S1. Acquire the initial image sequence of the construction site through image acquisition equipment deployed in a coordinated air-ground manner, and identify the construction process nodes in real time based on the initial image sequence; S2. For each frame image in the initial image sequence, extract the feature anchor points corresponding to the static ground features that remain unchanged within the frame; S3. Perform rigid alignment based on feature anchor point matching to generate a standardized image sequence with spatial alignment throughout the entire construction cycle; S4. Based on standardized image sequences, delineate dynamic construction monitoring target areas that match the current process node, extract grayscale gradient feature vectors and edge texture feature vectors within the target area, and synchronously transmit the extracted feature vectors to the control terminal via an air-to-ground communication link. S5. Based on the received feature vector, the control terminal retrieves the pre-stored standardized construction feature benchmarks for the corresponding process node, compares the current construction feature vector with the standardized construction feature benchmarks, and completes an accurate judgment on safe operation.
2. The power transmission and distribution construction monitoring method based on air-to-ground communication according to claim 1, characterized in that, The implementation of real-time identification of construction process nodes based on the initial image sequence in step S1 includes: S11. For each frame of the initial image sequence, extract the personnel operation vector and the device status vector from each frame. S12. Pre-construct a two-vector prior logic graph of power transmission and distribution construction procedures. The two-vector prior logic graph includes at least the personnel operation reference vector and equipment status reference vector corresponding to each standard construction procedure, as well as the sequential dependence of each procedure, the shortest / longest duration constraint of a single procedure, and the precondition constraint for starting the procedure. S13. Perform similarity matching between the personnel operation vector and the personnel operation reference vector of each process in the dual-vector prior logic graph, and output the first preliminary process result and the first confidence level; at the same time, perform similarity matching between the equipment state vector and the equipment state reference vector of the corresponding process in the dual-vector prior logic graph, and output the second preliminary process result and the second confidence level. S14. By using an attention mechanism to weightedly fuse the first and second confidence levels, the process with the highest overall matching degree is output as the final process node.
3. The power transmission and distribution construction monitoring method based on air-to-ground communication according to claim 1, characterized in that, Before step S2, image denoising needs to be performed on each frame of the initial image sequence, including: S211. For each frame of image, the local pixel range and the minimum eigenvalue of the first-order structure tensor are weighted and fused to calculate the pixel-level noise variance. S212. Construct a spatial structure tensor and a frequency domain amplitude-phase decoupling tensor respectively. The spatial structure tensor includes a first-order structure tensor and a second-order structure tensor. S213. Perform a two-dimensional Fourier transform on the image to decouple the amplitude and phase to obtain the phase tensor and the amplitude tensor. Keep the phase tensor unchanged, use the noise variance to correct the amplitude tensor, and combine the corrected amplitude tensor with the phase tensor to obtain the frequency domain denoised image through inverse Fourier transform. S214. Using the eigenvalue vectors of the first-order structure tensor and the second-order structure tensor as the basis, calculate the weights of the non-local image blocks and perform weighted filtering to obtain the spatial domain denoised image. S215. Based on the eigenvalues of the first-order structure tensor, an adaptive fusion weight is constructed. The weights are dynamically allocated according to the local structural features of the construction image. The frequency domain and spatial domain denoising results are fused to obtain the denoised image.
4. The power transmission and distribution construction monitoring method based on air-to-ground communication according to claim 3, characterized in that, Step S214, which uses the eigenvectors of the first-order and second-order structure tensors as similarity criteria, calculates the weights of non-local image patches and performs weighted filtering to obtain a spatially denoised image, includes the following: S2141. First, calculate the tensor similarity between any two image patches within the nonlocal search window; S2142. Normalize the calculated similarity to obtain the weighting coefficients for each image block, ensuring that the sum of the weights is 1. S2143. Based on the weighting coefficients, perform weighted fusion on all image blocks within the non-local search window to obtain a spatially structured denoised image.
5. The power transmission and distribution construction monitoring method based on air-to-ground communication according to claim 1, characterized in that, Step S2, which extracts feature anchor points corresponding to static features that remain unchanged within each frame of the initial image sequence, includes the following steps: S221. For each frame of the initial image sequence, first identify the preset static ground object categories in the power transmission and distribution construction scene through a lightweight semantic segmentation network, and generate the corresponding semantic mask. S222. Based on the geometric prior rules of static features, the semantic mask is filtered a second time to remove the falsely detected dynamic interference areas and obtain the static feature candidate areas. S223. Perform edge detection on each static ground feature candidate region and then extract the geometric vertex set of the candidate region through a polygon fitting algorithm; S224. Perform corner detection algorithm on each static ground feature candidate area to extract the corner set; S225. Perform spatial distance clustering on the geometric vertex set and corner set, and use the cluster centers where the spatial distance between the two types of feature points is less than a preset threshold as feature anchor points.
6. The power transmission and distribution construction monitoring method based on air-to-ground communication according to claim 1, characterized in that, The implementation of step S3, which performs rigid alignment based on feature anchor point matching to generate a standardized image sequence with spatial alignment throughout the entire construction cycle, includes: S31. Using the first frame after denoising as the reference frame, establish an initial set of anchor points corresponding to the reference frame and the current frame to be aligned based on feature anchor points. S32. Perform triangulation on the anchor point set in the reference frame to generate a topological triangle network. For each topological triangle, calculate its topological invariant vector in the reference frame. For the corresponding triangle in the current frame to be aligned, calculate its topological invariant vector and calculate the Euclidean distance between them. If the Euclidean distance is greater than the set topological error threshold, it is determined that there is an abnormal match of the anchor points in the triangle. The corresponding anchor point pair is removed to generate the final reliable matching set. S33. According to the category of the feature anchor points, the reliable matching set is divided into L subsets. For each subset, the random sampling consensus algorithm combined with the least squares method is used to estimate the local homography matrix corresponding to the land cover of that category, and the mean value of the anchor point reprojection error of that subset is calculated. S34. Calculate the adaptive weights of each local homography matrix based on the mean reprojection error, and fuse all local homography matrices according to the adaptive weights to obtain the global rigid alignment homography matrix. S35. Perform a complete perspective transformation resampling using a global rigid alignment homography matrix to obtain the pixel coordinates of the aligned image and complete the standardized image sequence.
7. The power transmission and distribution construction monitoring method based on air-to-ground communication according to claim 1, characterized in that, The implementation of step S4, which extracts the gray-level gradient feature vector and edge texture feature vector within the target region, includes: S41. Gray-level gradient feature vector extraction: For the defined target region ROI, a local pixel coordinate system with the associated feature anchor point as the origin is constructed. Multi-directional gray-level gradient calculation and normalization are performed on each pixel in the ROI to generate a global gray-level gradient feature vector. S42. For the defined target region ROI, calculate multi-scale texture features with edge pixels as the core, and generate an edge texture feature vector that matches the dimension of the gray-level gradient feature vector. S43. Based on the temporal continuity of the standardized image sequence, weighted moving average fusion is performed on the gray-level gradient feature vector and edge texture feature vector of the same ROI in three consecutive frames to remove feature mutations caused by noise and obtain the final output feature vector.
8. The power transmission and distribution construction monitoring method based on air-to-ground communication according to claim 1, characterized in that, In step S5, the feature vector of the current construction is compared with the standardized construction feature benchmark by weighted cosine similarity calculation.
9. A power transmission and distribution construction monitoring system based on air-to-ground communication, characterized in that, The system is applicable to the method described in any one of claims 1-8, comprising: an image acquisition and recognition module, a feature anchor point extraction module, a standardized image sequence module, a feature extraction and transmission module, and an intelligent judgment module; The image acquisition and recognition module acquires the initial image sequence of the construction site through image acquisition equipment deployed in a coordinated air-ground manner, and identifies the construction process nodes in real time based on the initial image sequence; The feature anchor point extraction module extracts the feature anchor points corresponding to the static ground features that remain unchanged within each frame of the initial image sequence. The standardized image sequence module performs rigid alignment based on feature anchor point matching to generate a standardized image sequence with spatial alignment throughout the entire construction cycle. The feature extraction and transmission module is based on a standardized image sequence, delineates a dynamic construction monitoring target area that matches the current process node, extracts the gray-level gradient feature vector and edge texture feature vector within the target area, and synchronously transmits the extracted feature vector to the control terminal through an air-to-ground communication link. The intelligent judgment module is a control terminal that, based on the received feature vector, retrieves the pre-stored standardized construction feature benchmarks for the corresponding process node, compares the current construction feature vector with the standardized construction feature benchmarks, and completes an accurate judgment on safe operation.