Three-dimensional reconstruction method and system for power transmission lines based on monocular channel monitoring camera

Abstract

The present invention discloses a method and system for 3D reconstruction of power transmission lines based on a monocular channel monitoring camera, the method comprising obtaining compensated image data through adaptive illumination compensation processing, obtaining directional feature data based on frequency domain feature analysis, constructing a texture enhancement model using graph theory to generate enhanced image data, and outputting feature aggregation data through temporal feature aggregation; generating 3D contour data by combining multi-conductor topology analysis and geometric constraint optimization; obtaining optimized depth data by using deep probability modeling and multi-scale consistency processing; and outputting system state data, 3D point cloud data and deformation parameter data through physical model construction and spatiotemporal constraint fusion. The present invention realizes high-precision 3D reconstruction of power transmission lines, and has strong environmental adaptability and computational efficiency.

Description

Technical Field

[0001] The present invention belongs to the field of computer vision and image processing, and in particular to a method and system for three-dimensional reconstruction of a power transmission line based on a monocular channel monitoring camera. Background Art

[0002] In the maintenance and management of power infrastructure, transmission lines are the neural network of the power system. With the rapid expansion of UHV transmission networks and the in-depth advancement of smart grid construction, traditional transmission line monitoring methods can no longer meet the needs of contemporary power grid operation and maintenance. Especially in scenarios such as complex terrain and long-distance power transmission, accurately obtaining three-dimensional spatial information of transmission lines is of great significance for preventing line failures, evaluating conductor status, and optimizing maintenance strategies. With its advantages of low equipment cost, convenient installation and maintenance, and real-time information acquisition, monocular cameras have shown great application potential in the field of transmission line monitoring, providing a new technical path for realizing intelligent and automated monitoring of transmission lines.

[0003] At present, the 3D reconstruction technology of power transmission lines mainly includes methods based on laser radar, multi-camera vision and monocular vision. Although laser radar has high measurement accuracy, the equipment is expensive and easily affected by weather; the multi-camera vision system requires strict camera calibration and synchronization, and has the problem of difficult deployment and maintenance in practical applications. The monocular vision method has attracted widespread attention due to its simple and economical characteristics. Existing research mainly focuses on model matching, motion recovery structure (SfM) and deep learning. The model matching method performs parameter fitting through a predefined conductor geometry model, but the effect is poor when dealing with complex deformations; the SfM method requires sufficient viewpoint changes to achieve reliable reconstruction, which is difficult to meet in fixed monitoring scenarios; although the deep learning-based method performs well in specific scenarios, it is limited by the acquisition of training data and the lack of generalization ability.

[0004] The current monocular 3D reconstruction technology of power transmission lines still faces many challenges: first, the wire appears as a slender target in the image, and its texture features are not significant. Under complex backgrounds and uneven lighting conditions, feature extraction and matching are prone to false detection and missed detection; second, due to the high geometric symmetry of the wire, it is difficult for the traditional feature point matching method to establish a reliable correspondence, especially when the wire is twisted or deformed significantly, the reconstruction accuracy is significantly reduced; third, the image quality fluctuations caused by environmental factors (such as atmospheric disturbances and lighting changes) make the reconstruction results based on single-frame images unstable and difficult to ensure time consistency; fourth, there is a lack of uncertainty quantification and reliability assessment mechanism for the reconstruction results, which makes it difficult to provide reliable confidence indicators for subsequent state assessment and early warning decisions; fifth, existing methods often ignore the physical characteristics and constraints of the wire system, resulting in the reconstruction results that may violate physical laws, affecting the practicality of the results. These technical difficulties seriously restrict the application of monocular 3D reconstruction technology of power transmission lines in actual engineering. Summary of the invention

[0005] The purpose of the invention is to provide a method and system for three-dimensional reconstruction of power transmission lines based on a monocular channel monitoring camera to solve the above-mentioned problems existing in the prior art.

[0006] The technical solution is a three-dimensional reconstruction method of a power transmission line based on a monocular channel monitoring camera, comprising the following steps:

[0007] S1. Obtain original camera image data, generate compensated image data through adaptive illumination compensation processing; perform frequency domain feature analysis on the compensated image data to obtain directional feature data; construct a texture enhancement model based on the directional feature data using graph theory methods to generate enhanced image data; perform temporal feature aggregation processing based on the enhanced image data to output feature aggregation data;

[0008] S2. Generate topological constraint data through multi-conductor topological analysis based on feature aggregation data and directional feature data; extract geometric constraint conditions based on directional feature data and feature aggregation data; construct first constraint energy data based on topological constraint data and geometric constraint conditions, perform time-deformation compensation processing, and obtain deformation compensation data; generate optimized constraint data through constraint optimization calculation based on deformation compensation data; perform robustness verification based on optimized constraint data, and output three-dimensional contour data;

[0009] S3. Based on the three-dimensional contour data, generate depth probability data through depth probability modeling; based on the depth probability data, perform feature matching optimization to obtain feature depth data; based on the feature depth data, perform multi-scale consistency processing to generate consistent depth data; based on the consistent depth data and pre-stored camera calibration parameters, perform geometric constraint optimization to obtain constrained depth data; based on the constrained depth data, output optimized depth data, depth uncertainty data and depth consistency data through dynamic update processing;

[0010] S4. Based on the optimized depth data and pre-stored historical environmental parameters, physical constraint data is constructed through the pre-configured physical model. Based on the physical constraint data and pre-stored physical model parameters, dynamic tension data is obtained through tension deformation analysis. Based on the optimized depth data and depth consistency data, spatiotemporal constraint data is generated through spatiotemporal consistency processing. Based on the dynamic tension data, spatiotemporal constraint data and depth uncertainty data, physical constraint fusion optimization is performed to finally output system state data, three-dimensional point cloud data and deformation parameter data.

[0011] The 3D reconstruction system of power transmission lines based on a monocular channel monitoring camera includes:

[0012] at least one processor; and,

[0013] a memory communicatively connected to at least one of the processors; wherein,

[0014] The memory stores instructions that can be executed by the processor, and the instructions are used to be executed by the processor to implement the three-dimensional reconstruction method of the power transmission line based on the monocular channel monitoring camera.

[0015] Beneficial effects: The present invention overcomes imaging interference in complex environments, ensures the accuracy and physical rationality of reconstruction results, provides reliable depth information and confidence assessment, and realizes accurate analysis of the conductor system status; it can adapt to complex on-site environments, provide reliable monitoring results, and realize high-precision three-dimensional reconstruction and status monitoring of transmission lines based on a monocular camera. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 Flow chart of the method of the present invention.

[0017] Figure 2 This is a flow chart of step S1 of the present invention.

[0018] Figure 3 This is a flow chart of step S2 of the present invention.

[0019] Figure 4 This is a flow chart of step S3 of the present invention.

[0020] Figure 5 This is a flow chart of step S4 of the present invention. DETAILED DESCRIPTION

[0021] like Figure 1 As shown, the present application proposes a method for three-dimensional reconstruction of a power transmission line based on a monocular channel monitoring camera, comprising the following steps:

[0022] S1. Obtain original camera image data, generate compensated image data through adaptive illumination compensation processing; perform frequency domain feature analysis on the compensated image data to obtain directional feature data; construct a texture enhancement model based on the directional feature data using graph theory methods to generate enhanced image data; perform temporal feature aggregation processing based on the enhanced image data to output feature aggregation data;

[0023] S2. Generate topological constraint data through multi-conductor topological analysis based on feature aggregation data and directional feature data; extract geometric constraint conditions based on directional feature data and feature aggregation data; construct first constraint energy data based on topological constraint data and geometric constraint conditions, perform time-deformation compensation processing, and obtain deformation compensation data; generate optimized constraint data through constraint optimization calculation based on deformation compensation data; perform robustness verification based on optimized constraint data, and output three-dimensional contour data;

[0024] S3. Based on the three-dimensional contour data, generate depth probability data through depth probability modeling; based on the depth probability data, perform feature matching optimization to obtain feature depth data; based on the feature depth data, perform multi-scale consistency processing to generate consistent depth data; based on the consistent depth data and pre-stored camera calibration parameters, perform geometric constraint optimization to obtain constrained depth data; based on the constrained depth data, output optimized depth data, depth uncertainty data and depth consistency data through dynamic update processing;

[0025] S4. Based on the optimized depth data and pre-stored historical environmental parameters, physical constraint data is constructed through the pre-configured physical model. Based on the physical constraint data and pre-stored physical model parameters, dynamic tension data is obtained through tension deformation analysis. Based on the optimized depth data and depth consistency data, spatiotemporal constraint data is generated through spatiotemporal consistency processing. Based on the dynamic tension data, spatiotemporal constraint data and depth uncertainty data, physical constraint fusion optimization is performed to finally output system state data, three-dimensional point cloud data and deformation parameter data.

[0026] like Figure 2 As shown, according to one aspect of the present application, step S1 is further:

[0027] S11, obtaining original camera image data, calculating image brightness distribution, and obtaining brightness histogram data; performing local area entropy analysis on the brightness histogram data to generate local entropy weight data; calculating a compensation coefficient based on the local entropy weight data; performing pixel-level multiplication operation on the compensation coefficient and the original camera image data, and outputting compensated image data;

[0028] S12, based on the compensated image data, performing a two-dimensional Fourier transform to obtain frequency domain feature data; constructing a predetermined number of directional filters to generate a directional filter group; performing a convolution operation on the frequency domain feature data and the directional filter group to obtain directional response data; performing a maximum value extraction on the directional response data and outputting the directional feature data;

[0029] S13, based on the compensated image data and the directional feature data, construct a pixel association graph to generate graph structure data; calculate the edge weights in the graph structure data to obtain weight matrix data; perform eigenvalue decomposition on the weight matrix data to obtain eigenvector data; based on the eigenvector data, reconstruct the compensated image data and output enhanced image data;

[0030] S14. Based on the enhanced image data, obtain the enhanced image data sequence within the preset time window and construct a time series feature tensor; calculate the feature similarity of adjacent moments in the time series feature tensor and generate time series weight data; perform weighted combination of the time series weight data and the time series feature tensor and output feature aggregation data.

[0031] In one embodiment of the present application, the local entropy weight compensation coefficient C(x, y) = α·exp(-H(x, y) / H_max)·(1 + β·G(x, y))·(1 - γ·E(x, y)) is calculated; wherein H(x, y) = -∑p_i·log(p_i) is the local entropy value, p_i is the local histogram probability distribution; H_max is the maximum entropy value of the image; G(x, y) = ||▽I(x, y)|| is the image gradient amplitude; E(x, y) = exp(-||x-x_c|| 2 / (2σ 2 )) is the Gaussian attenuation term; α is the global gain coefficient, ranging from [0.5, 2]; β is the gradient enhancement coefficient, ranging from [0, 1]; γ is the edge preservation coefficient, ranging from [0, 1]; x_c is the image center coordinate; σ is the Gaussian kernel parameter.

[0032] The edge weight is constructed using a texture enhancement optimization algorithm based on graph theory: W(i, j) = exp(-||F(i)-F(j)|| 2 / σf 2 )·exp(-||P(i)-P(j)|| 2 / σp 2)·(1 + μ·C(i,j))·exp(-η·R(i,j)); where F(i) is the feature vector of pixel i; P(i) is the spatial position of pixel i; C(i,j) = cos(v(i), v(j)) is the directional consistency term; R(i,j) = ||I(i)-I(j)|| 2 / σr 2 is the brightness similarity; v(i) is the direction vector at pixel i; σf is the feature scale parameter; σp is the spatial scale parameter; σr is the brightness scale parameter; μ is the direction weight coefficient, ranging from [0, 2]; η is the brightness weight coefficient, ranging from [0.1, 1]; W(i, j) is the edge weight between pixels i and j; I(i) is the grayscale value of pixel i.

[0033] This embodiment realizes the conversion from monocular camera images to high-quality feature aggregation data through a multi-level image processing and feature extraction process. First, the imaging problem of the transmission line in the complex lighting environment is solved by adaptive illumination compensation processing, and the uneven image quality caused by weather, time and other factors is effectively compensated. Secondly, the frequency domain feature analysis method is used to extract directional features. Compared with the traditional spatial domain analysis method, it can better capture the global direction information of the wire and has stronger robustness to noise. Through the texture enhancement model based on graph theory, the topological relationship between pixels is established, and the global optimization characteristics of the graph are used to enhance the texture features of the wire, thereby improving the accuracy of subsequent three-dimensional reconstruction. Finally, the temporal feature aggregation processing effectively suppresses the random errors caused by camera shake, atmospheric disturbances, etc. by weighted fusion of the features of continuous frames. This embodiment not only improves the accuracy and robustness of feature extraction, but also realizes the temporal consistency of features through temporal aggregation, providing high-quality input data for subsequent three-dimensional reconstruction.

[0034] According to one aspect of the present application, step S11 is further:

[0035] S111, obtaining original camera image data, and converting it into HSV color space to obtain HSV image data; extracting V channel data in the HSV image data to generate brightness channel data; performing histogram statistical calculation on the brightness channel data, and outputting brightness histogram data;

[0036] S112, based on the brightness histogram data, using an adaptive window to perform block processing on the image to obtain regional block data; calculating the grayscale distribution probability of each sub-region in the regional block data to generate probability distribution data; performing local entropy calculation on the probability distribution data to output local entropy weight data;

[0037] S113, based on the local entropy weight data, construct a weight mapping function based on the local entropy to obtain weight mapping data; perform Gaussian smoothing on the weight mapping data to generate smoothed weight data; based on the smoothed weight data, calculate the compensation coefficient of each pixel position, and output the compensation coefficient data;

[0038] S114, based on the compensation coefficient data and the original camera image data, respectively mapping the compensation coefficients for the RGB channels to obtain channel compensation data; merging the channel compensation data to generate fused compensation data; performing edge preserving filtering on the fused compensation data to output compensated image data.

[0039] In one embodiment of the present application, the HSV color space adaptive conversion method: C_hsv(x, y) = [h(x, y), s(x, y), v'(x, y)]; wherein v'(x, y) = v(x, y)·(1 + α·ΔL(x, y))·exp(-β·E(x, y)); ΔL(x, y) = L_avg - L(x, y) is the brightness difference; L_avg is the local average brightness; E(x, y) = ||▽v(x, y)|| 2 is the gradient energy; h(x, y) is the hue component; s(x, y) is the saturation component; v(x, y) is the brightness component; α is the brightness compensation coefficient, ranging from [0.5, 2]; β is the edge preservation coefficient, ranging from [0.1, 1]; C_hsv(x, y) is the HSV color vector at the coordinate point (x, y); ▽v(x, y) is the gradient operator of the brightness component.

[0040] Regional block adaptive window algorithm: W(x, y) = W_base·(1 + λ·S(x, y))·exp(-μ·D(x, y)); where W_base is the base window size; S(x, y) = σ(x, y) / σ_max is the local variance ratio; D(x, y) = ||[x, y] - [x_c, y_c]|| 2 / r 2 is the distance term; σ(x, y) is the local standard deviation; σ_max is the maximum standard deviation of the entire image; [x_c, y_c] is the coordinate of the image center; r is the reference radius; λ is the variance weight coefficient, ranging from [0, 2]; μ is the distance attenuation coefficient, ranging from [0.1, 1]; W(x, y) is the adaptive window size at the coordinate point (x, y).

[0041] Local entropy weight mapping function: M(H) = α·(1 - exp(-H / H_max))·(1 + β·G(H))·exp(-γ·V(H)); where H is the local entropy value; H_max is the maximum entropy value; G(H) = ||▽H||2 is the entropy gradient; V(H) =var(H) is the local entropy variance; α is the global mapping coefficient, ranging from [0.5, 2]; β is the gradient gain coefficient, ranging from [0, 1]; γ is the variance suppression coefficient, ranging from [0.1, 1]; M(H) is the mapping function from entropy value to weight; ▽H is the entropy gradient operator.

[0042] This embodiment effectively solves the quality problem of transmission line images under complex lighting environments through adaptive illumination compensation processing. First, through HSV color space conversion and brightness channel extraction, the brightness information of the image is better separated, avoiding the interference caused by the correlation of the three channels in the RGB space. The regional block processing of the adaptive window fully considers the local characteristics of the image, so that the compensation effect can adapt to the illumination changes in different regions. Through the calculation of local entropy and weight mapping, an adaptive compensation mechanism based on information entropy is established, which can effectively enhance the visibility of low-illuminance areas while maintaining image details. Gaussian smoothing ensures the spatial continuity of the compensation coefficient and avoids the block effect during the compensation process. The final edge-preserving filter not only ensures the smoothness of the compensated image, but also well maintains the sharp features of the wire edge. This embodiment not only improves the visual quality of the image, but also provides high-quality input data for subsequent feature extraction.

[0043] According to one aspect of the present application, step S12 is further:

[0044] S121, reading the compensated image data, using the Hanning window function to perform edge transition processing to obtain window function data; performing fast Fourier transform on the window function data to generate fast Fourier transform (FFT) data; performing centering and amplitude normalization processing on the FFT data, and outputting frequency domain feature data;

[0045] S122, calculating the direction sampling points according to the preset angle interval to obtain direction sampling data; constructing a direction selective filter based on the Log-Gabor filter basis function to generate Gabor filter data; combining the Gabor filter data with an anisotropic Gaussian kernel to obtain kernel function data; performing frequency domain transformation processing on the kernel function data, and outputting a direction filter group;

[0046] S123, obtaining frequency domain feature data and a directional filter group, and using a fast convolution algorithm to calculate convolution result data; performing phase consistency analysis on the convolution result data to generate phase feature data; fusing the phase feature data with the amplitude information, and outputting directional response data;

[0047] S124, receiving directional response data, using a non-maximum suppression algorithm to perform local maximum detection to obtain peak data; performing sub-pixel precision optimization on the peak data to generate optimized peak data; calculating the main direction and secondary direction features in combination with the optimized peak data, and outputting directional feature data.

[0048] In one embodiment of the present application, the Hanning window function edge transition processing is: W(r)=0.5·(1 + cos(πr / R))·(1 + α·G(r))·exp(-β·E(r)); where r is the distance from the pixel to the center of the window; R is the window radius; G(r) = exp(-r 2 / 2σ 2 ) is a Gaussian weighted term; E(r) = ||▽I(r)|| 2 is the edge strength; α is the Gaussian weight coefficient, ranging from [0, 1]; β is the edge preservation coefficient, ranging from [0.1, 2]; σ is the Gaussian kernel parameter; W(r) is the window weight at a distance r from the center; I(r) is the image intensity at a distance r.

[0049] Directional sampling point optimization algorithm: θ(k) =θ_base·(1 + λ·R(k))·exp(-μ·D(k)); where θ_base is the base sampling angle; R(k) = res(k) / res_max is the response intensity ratio; D(k) = ||θ(k) -θ_pred|| 2 is the direction deviation; res(k) is the response value of direction k; res_max is the maximum response value; θ_pred is the predicted direction; λ is the response weight coefficient, ranging from [0, 2]; μ is the deviation penalty coefficient, ranging from [0.1, 1]; θ(k) is the sampling angle of the kth direction.

[0050] Directional Log-Gabor filter design algorithm: G(ω,θ) = exp(-log(ω / ω0) 2 / (2log(σω) 2 ))·exp(-(θ-θ0) 2 / (2σθ 2))·(1+λ·D(ω))·exp(-α·|H(θ)|); where ω is the radial frequency; θ is the directional angle; ω0 is the center frequency; θ0 is the reference direction; σω is the radial bandwidth parameter; σθ is the angular bandwidth parameter; D(ω) = 1 -exp(-β·|ω-ω0|) is the frequency modulation term; H(θ) = cos(n·(θ-θ0)) is the directional harmonic function; λ is the frequency selectivity coefficient, ranging from [0, 1]; α is the directional sensitivity coefficient, ranging from [0.1, 2]; β is the frequency modulation coefficient, ranging from [0.5, 5]; n is the harmonic order, usually 2 or 4; G(ω, θ) is the response function of the filter in the frequency domain.

[0051] Phase consistency analysis method: PC(x, y) = (∑A_n·cos(φ_n - φ_mean)) / (∑A_n +ε)·(1 + α·E(x, y))·exp(-β·V(φ)); where A_n is the amplitude; φ_n is the phase angle; φ_mean is the average phase; E(x, y) = ||▽I(x, y)|| 2 is the gradient energy; V(φ) = var(φ_n) is the phase variance; ε is a small positive number to prevent division by zero; α is the gradient weight coefficient, ranging from [0, 1]; β is the variance penalty coefficient, ranging from [0.1, 2]; PC(x, y) is the phase consistency score at the coordinate point (x, y); I(x, y) is the image intensity.

[0052] This embodiment combines frequency domain analysis with directional feature extraction to achieve accurate extraction of wire directional features. Edge transition processing is performed through the Hanning window function, which effectively suppresses the spectrum leakage effect in frequency domain transformation. Fast Fourier transform combined with centralization and normalization processing not only improves the computational efficiency, but also enhances the scale invariance of feature extraction. The design of the directional filter group adopts the Log-Gabor basis function, which not only ensures directional selectivity but also has good frequency band characteristics. By combining the Gabor filter with the anisotropic Gaussian kernel, the response capability to specific directional features is further enhanced. The introduction of phase consistency analysis effectively extracts important structural information, while non-maximum suppression and sub-pixel optimization ensure the precise positioning of the feature position. This embodiment can not only accurately extract the directional features of the wire, but also has strong robustness to image noise and illumination changes.

[0053] According to one aspect of the present application, step S13 is further:

[0054] S131, reading the compensated image data and the directional feature data, using an adaptive neighborhood partitioning algorithm to establish a pixel connection relationship to obtain connection relationship data; performing a non-local similarity calculation on the connection relationship data to generate similarity data; constructing a sparse adjacency matrix based on the similarity data, and outputting graph structure data;

[0055] S132, obtaining graph structure data, using anisotropic diffusion model to calculate edge weights to obtain diffusion weight data; performing structural tensor analysis on the diffusion weight data to generate tensor feature data; optimizing the tensor feature data in combination with local directional consistency constraints, and outputting weight matrix data;

[0056] S133, receiving weight matrix data, using Krylov subspace iteration method to perform large-scale eigenvalue calculation to obtain eigenvalue data; performing spectral analysis on the eigenvalue data to generate spectral distribution data; selecting the optimal eigenvector based on the spectral distribution data, and outputting the eigenvector data;

[0057] S134, reading feature vector data and compensated image data, constructing a reconstruction model based on manifold learning to obtain reconstructed model data; iteratively optimizing the reconstructed model data to generate optimized result data; refining the optimized result data in combination with edge preservation constraints, and outputting enhanced image data.

[0058] This embodiment achieves the enhancement of texture features through graph theory methods and improves the feature expression ability of wire images. The pixel connection relationship established by the adaptive neighborhood partitioning algorithm fully considers the local structural features of the image, so that the constructed graph structure can better reflect the texture characteristics of the wire. The introduction of the anisotropic diffusion model enables the edge weight calculation to adapt to texture features in different directions, especially in important feature areas such as wire edges. Structural tensor analysis further enhances the directional sensitivity of weight calculation and ensures the directional consistency of texture enhancement. Large-scale eigenvalue calculation is performed through the Krylov subspace iteration method to improve the computational efficiency, while spectral analysis helps to select the most representative eigenvectors. The reconstruction model based on manifold learning organizes local features into a continuous manifold structure to ensure the global consistency of the enhanced image. This embodiment not only enhances the texture features of the wire, but also maintains the integrity of the edge information.

[0059] like Figure 3 As shown, according to one aspect of the present application, step S2 is further:

[0060] S21. Based on the feature aggregation data and the directional feature data, construct a wire spatial parameter equation to obtain spatial parameter data; based on the spatial parameter data, calculate the relative position and angle relationship between the wires to generate position relationship data; extract topological features from the position relationship data, calculate topological invariants, and output topological constraint data;

[0061] S22, constructing a local coordinate system based on the directional feature data; projecting the feature aggregation data into the local coordinate system to generate a geometric constraint matrix; calculating a first constraint energy function based on the geometric constraint matrix and the topological constraint data, and outputting first constraint energy data;

[0062] S23, constructing a wire deformation state space based on the first constraint energy data to generate state space data; calculating a deformation prediction matrix based on the state space data to obtain predicted deformation data; comparing the predicted deformation data with the actual observed value, calculating a deformation compensation term, and outputting the deformation compensation data;

[0063] S24, constructing a global optimization objective function based on the first constraint energy data, the deformation compensation data and the predefined regularization term to obtain objective function data; calculating the gradient and quadratic approximation of the objective function data to generate optimization parameter data; using the trust region method to iteratively solve the optimization parameter data and output optimization constraint data;

[0064] S25. Calculate the structural stability index of the optimized constraint data to generate stability data; construct a confidence map based on the stability data to obtain confidence data; perform threshold processing on the confidence data to identify unstable areas; generate three-dimensional contour data based on the unstable areas and the optimized constraint data.

[0065] In one embodiment of the present application, a multi-conductor topology invariant extraction algorithm is: T(L)=∑(K(s)·τ(s))ds+ α·∮(dθ / ds) 2 ds + β·∫∫(Ξ 2 z / Ξx 2 +Ξ 2 z / Ξy 2 ) 2 dxdy; where K(s) is the curvature of the curve; τ(s) is the torsion of the curve; Ξ is the partial derivative, dθ / ds is the rate of change of the tangential angle; Ξ 2 z / Ξx 2 and 2 z / Ξy 2 is the second-order partial derivative of the surface; s is the curve parameter; α is the curvature weight coefficient, ranging from [0.1, 1]; β is the surface smoothness coefficient, ranging from [0.01, 0.5]; T(L) is the topological invariant; L is the wire space curve; z is the surface height function; x, y are plane coordinates.

[0066] Deformation state prediction and compensation algorithm: X(t+1) =Φ(t)·X(t) + ω·B(t)·u(t) + γ·Q(t)·exp(-λ·||e(t)|| 2) where X(t) is the state vector; Φ(t) is the state transfer matrix; B(t) is the control matrix; u(t) is the control input; Q(t) is the process noise matrix; e(t) is the prediction error vector; ω is the control gain coefficient, ranging from [0.1, 2]; γ is the adaptive correction coefficient, ranging from [0, 1]; λ is the error attenuation coefficient, ranging from [0.5, 5].

[0067] Trust region constrained optimization algorithm: min f(x) +ρ·||D(x)|| 2 +μ·R(x), st ||x-x_k||≤Δ_k; where f(x) is the objective function; D(x) is the constraint violation measure; R(x) is the regularization term; x_k is the current iteration point; Δ_k is the trust region radius; ρ is the constraint penalty coefficient, range [1, 100]; μ is the regularization coefficient, range [0.01, 1]; Δ_k update rule: Δ_(k+1) = β·Δ_k·(1+α·r_k), r_k is the ratio of the actual reduction to the predicted reduction; α is the update gain, range [0.1, 2]; β is the shrinkage coefficient, range [0.5, 1]; r_k is the trust region radius update ratio.

[0068] This embodiment combines multi-conductor topology analysis with geometric constraint optimization to achieve accurate reconstruction of the spatial structure of the transmission line. By analyzing the relative position and angle relationship between the conductors, a constraint system based on topological invariants is established, which has strong adaptability to conductor deformation. Especially when dealing with scenes where multiple conductors cross or run in parallel, topological constraints can effectively maintain the spatial relationship between the conductors, avoiding the common spatial structure confusion problem in traditional reconstruction methods. The constraint energy function constructed in combination with geometric constraints effectively suppresses the discontinuity and jump phenomenon in the reconstruction results by considering the curvature continuity and spatial smoothness of the conductors. The time-deformation compensation process introduces a state space model, which improves the reconstruction accuracy under severe weather conditions such as strong winds by predicting and compensating for the dynamic deformation of the conductors. Finally, the reliability and stability of the reconstruction results are ensured through constraint optimization calculation and robustness verification. This embodiment not only ensures the geometric accuracy of the reconstruction results, but also ensures the physical rationality of the results through multiple constraints.

[0069] According to one aspect of the present application, step S21 is further:

[0070] S211, receiving feature aggregation data and directional feature data, constructing a three-dimensional space curve using spline interpolation to obtain curve parameter data; performing differential geometry analysis on the curve parameter data to generate curvature torsion data; establishing a Frenet frame in combination with the curvature torsion data, and outputting the space parameter data;

[0071] S212, obtaining spatial parameter data, using quaternion representation to calculate spatial rotation to obtain rotation matrix data; performing coordinate transformation on the rotation matrix data to generate transformation matrix data; calculating the relative configuration between the wires based on the transformation matrix data, and outputting position relationship data;

[0072] S213, reading the position relationship data, extracting the topological features by using the continuous homology method to obtain homology feature data; performing Betti number calculation on the homology feature data to generate topological data; calculating the discrete curvature flow in combination with the topological data, and outputting the manifold feature data;

[0073] S214, receiving manifold feature data and positional relationship data, constructing a Reeb graph to describe the topological structure and obtaining topological graph data; calculating invariant moments of the topological graph data to generate invariant data; optimizing the invariant data in combination with geometric constraints, and outputting topological constraint data.

[0074] This embodiment achieves an accurate description of the spatial structure of the conductor by combining spatial parameterization and topological feature analysis. The three-dimensional space curve constructed by spline interpolation provides a continuous expression of the conductor shape, while differential geometry analysis provides important geometric features such as curvature and torsion. The establishment of the Frenet frame provides a reliable local coordinate system for subsequent spatial analysis. The introduction of quaternion representation simplifies the calculation of spatial rotation and avoids the universal lock problem in the Euler angle representation. The application of the continuous homology method extracts the topological features of the conductor system, which are invariant to deformation and provide a reliable reference for subsequent state monitoring. The construction of the Reeb graph further abstracts the topological structure of the conductor system, allowing the system to effectively handle complex spatial layouts. This embodiment not only ensures the accuracy of spatial feature extraction, but also maintains the stability of the topological structure.

[0075] According to one aspect of the present application, step S23 is further:

[0076] S231, based on the first constraint energy data, construct a generalized coordinate system through a preconfigured wire mechanics model to obtain coordinate system data; perform modal decomposition on the coordinate system data to generate modal characteristic data; combine the modal characteristic data with the preconfigured system dynamics equation to construct a state transfer equation, and output state space data;

[0077] S232, based on the state space data, using adaptive Kalman filtering to construct a predictor to obtain filter data; performing multi-step prediction on the filter data to generate multi-step prediction data; using Lyapunov stability analysis method to optimize the multi-step prediction data to obtain predicted deformation data;

[0078] S233, based on the predicted deformation data, construct a multi-scale decomposition based on wavelet transform to obtain scale decomposition data; perform singular value decomposition on the scale decomposition data to generate singular value data; based on the singular value data, extract the main deformation mode and output the deformation mode data;

[0079] S234. Based on the deformation pattern data and the predicted deformation data, the residual between the actual observed value and the predicted value is calculated to obtain the residual vector data; robust M-estimation is performed on the residual vector data to generate robust estimation data; based on the robust estimation data, an adaptive compensation function is constructed to output the deformation compensation data.

[0080] In one embodiment of the present application, the modal decomposition optimization algorithm: Φ(x, t) = ∑(a_i(t)·ψ_i(x))·exp(-λ·T(t))·(1 + μ·C(x)); where a_i(t) is the modal coefficient; ψ_i(x) is the modal function; T(t) = t / T_0 is the normalized time; C(x) = ||▽ 2 ψ(x)|| 2 is the curvature term; λ is the time attenuation coefficient, ranging from [0.1, 1]; μ is the curvature weight coefficient, ranging from [0, 2]; T_0 is the reference time scale; Φ(x, t) is the modal function value at the spacetime position (x, t); ▽ 2 ψ(x) is the Laplace operator of the mode function.

[0081] Wavelet transform multi-scale decomposition: W(s, τ) = ∫x(t)·ψ((t-τ) / s)dt·exp(-α·E(s))·(1 +β·R(τ)); where x(t) is the signal; ψ(t) is the wavelet basis function; s is the scale parameter; τ is the translation parameter; E(s) = ||ΞW / Ξs|| 2 is the scale energy; R(τ) = |W(s, τ)| 2 / E_avg is the relative energy; α is the energy weight coefficient, ranging from [0.1, 2]; β is the relative energy gain coefficient, ranging from [0, 1]; E_avg is the average energy; W(s, τ) is the wavelet coefficient at scale s and translation τ; ΞW / Ξs is the partial derivative of the wavelet coefficient with respect to the scale.

[0082] Robust M-estimation compensation method: ρ(e) = (|e| 2 / (σ 2 + |e| 2 ))·exp(-α·V(e))·(1 +β·G(e)); where e is the residual; σ is the scale parameter; V(e) = var(e) is the residual variance; G(e) = ||▽e|| 2is the residual gradient; α is the variance penalty coefficient, ranging from [0.1, 1]; β is the gradient gain coefficient, ranging from [0, 2]; ρ(e) is the robust loss function of the M-estimator; ▽e is the gradient of the residual.

[0083] This embodiment achieves accurate prediction and compensation of wire deformation through state space modeling and multi-level prediction analysis. The construction of the generalized coordinate system and modal decomposition provide a complete mathematical description for wire movement, especially modal decomposition can effectively separate the main deformation modes. The construction of the adaptive Kalman filter combines the system dynamics equation, which not only improves the accuracy of the prediction, but also can adaptively adjust the filter parameters to adapt to different working conditions. The introduction of Lyapunov stability analysis ensures the stability of the prediction system and avoids prediction divergence. The multi-scale decomposition of the wavelet transform enables the system to analyze deformation characteristics at different time scales, while the singular value decomposition helps to extract the most significant deformation mode. The application of Robust M-estimation improves the system's resistance to outliers, making deformation compensation more robust. This embodiment not only ensures the accuracy of deformation prediction, but also achieves effective handling of abnormal situations.

[0084] According to one aspect of the present application, step S24 is further:

[0085] S241, reading the constraint energy data and the deformation compensation data, constructing a sparse regularization term based on the L1 norm to obtain regularization term data; performing adaptive weight adjustment on the regularization term data to generate weight coefficient data; combining the weight coefficient data with the constraint term to construct an augmented Lagrangian function, and outputting the objective function data;

[0086] S242, obtaining objective function data, using automatic differentiation to calculate the first-order derivative to obtain gradient vector data; using the quasi-Newton method to estimate the Hessian matrix of the gradient vector data to generate second-order derivative data; analyzing the stability of the second-order derivative data based on Lipschitz continuity, and outputting optimization parameter data;

[0087] S243, receiving optimization parameter data, constructing a trust region subproblem based on the Dogleg method to obtain subproblem data; performing conjugate gradient solution on the subproblem data to generate step data; updating the step data in combination with the adaptive trust region radius, and outputting iterative data;

[0088] S244, read the iteration data and the objective function data, calculate the descent amount of the objective function to obtain the descent amount data; perform convergence analysis on the descent amount data to generate convergence index data; perform optimality verification on the convergence index data based on the KKT condition, and output the optimization constraint data.

[0089] This embodiment achieves the global optimal solution under constraints by constructing a complex optimization framework. The sparse regularization term of the L1 norm effectively controls the complexity of the solution and avoids the overfitting problem. The adaptive weight adjustment mechanism can dynamically adjust the influence of different constraints according to their importance, ensuring the rationality of the optimization process. Automatic differentiation calculation provides high-precision gradient information, and the application of the quasi-Newton method effectively reduces the complexity of the second-order derivative calculation. Lipschitz continuity analysis ensures the numerical stability of the optimization process. The trust region subproblem constructed by the Dogleg method provides a more flexible step size selection strategy, and the application of the conjugate gradient method improves the solution efficiency. The update strategy of the adaptive trust region radius ensures the rapid convergence of the optimization process. The verification of the KKT condition provides a theoretical guarantee for the reliability of the optimal solution. This embodiment not only ensures the accuracy of the solution, but also improves the computational efficiency.

[0090] like Figure 4 As shown, according to one aspect of the present application, step S3 is further:

[0091] S31, based on the three-dimensional contour data, calculating the initial depth distribution of each point to generate initial depth distribution data; based on the initial depth distribution data, extracting the depth relationship between adjacent points, constructing a depth consistency constraint, and obtaining depth constraint data; combining the initial depth distribution data with the depth constraint data, and outputting depth probability data;

[0092] S32, based on the depth probability data and the confidence data, extract feature points to obtain feature point data; calculate the matching relationship of the feature point data to generate feature matching data; based on the feature matching data and the confidence data, calculate the reliability weight of the feature points to obtain feature weight data; based on the feature matching data and the feature weight data, optimize the depth value and output the feature depth data;

[0093] S33, performing multi-scale decomposition on the feature depth data to generate depth pyramid data; calculating the depth propagation relationship between each level of the depth pyramid data to obtain the depth propagation data; constructing an optimization target, optimizing the depth propagation data, and outputting consistent depth data; wherein the optimization target includes a data item, a smoothing item, and a scale item;

[0094] S34, based on the pre-stored camera calibration parameters and consistency depth data, construct PnP constraints and generate PnP constraint data; based on the PnP constraint data, calculate the distance from the spatial point to the straight line, obtain the geometric consistency score, and generate geometric score data; merge the PnP constraint data with the geometric score data, update the depth probability distribution, and output the constraint depth data;

[0095] S35. Based on the confidence data, the depth update weight is calculated to obtain the updated weight data; based on the updated weight data, the constrained depth data is weightedly combined with the pre-stored historical depth data to generate depth status data; anomaly detection and correction is performed on the depth status data, and finally optimized depth data, depth uncertainty data and depth consistency data are output.

[0096] In one embodiment of the present application, the deep probability distribution modeling algorithm: P(d|I) = N(d|μ(x, y), σ 2 (x, y))·exp(-α·V(d))·(1 + β·C(d))·exp(-γ·E(d)); where d is the depth value; μ(x, y) is the depth expectation; σ 2 (x, y) is the depth variance; V(d) = ∑w_i·||d-d_i|| 2 is the spatial consistency term; C(d) = exp(-||▽d|| 2 / 2τ 2 ) is the depth continuity term; E(d) = ||d-d_prior|| 2 is the prior constraint; w_i is the spatial weight; d_i is the neighborhood depth value; τ is the smoothing parameter; α is the consistency coefficient, ranging from [0.1, 2]; β is the continuity coefficient, ranging from [0, 1]; γ is the prior weight, ranging from [0.01, 0.5]; P(d|I) is the posterior probability of depth d under the condition of a given image I; N(·) is the Gaussian distribution function; d_prior is the depth prior value; ▽d is the depth gradient operator.

[0097] Multi-scale deep consistency optimization algorithm: E(D) = ∑λ_l·||D_l - U(D_(l+1))|| 2 + μ·∑||L(D_l)|| 2 + η·∑Ψ(▽D_l); where D_l is the depth map of the lth layer; U(·) is the upsampling operator; L(·) is the Laplacian operator; Ψ(x) = sqrt(x 2 + ε 2 ) is the robust norm; ▽D_l is the depth gradient; λ_l is the inter-layer weight coefficient, λ_l = 2 -l ; μ is the smoothing coefficient, ranging from [0.1, 1]; η is the edge preservation coefficient, ranging from [0.01, 0.5]; ε is the norm parameter, with a value of 1e-6; E(D) is the depth consistency energy function; D is the depth image.

[0098] This embodiment achieves accurate estimation and uncertainty quantification of the three-dimensional depth information of the conductor through deep probability modeling and multi-level optimization strategies. First, a probability distribution model of the spatial position of the conductor is established through deep probability modeling, which can effectively handle the uncertainty and ambiguity in depth estimation. Feature matching optimization combines local descriptors and geometric constraints to improve the accuracy and robustness of feature point matching. Multi-scale consistency processing realizes consistency constraints on depth information at different spatial scales by establishing a deep propagation relationship, effectively suppressing outliers in local depth estimation. Geometric constraint optimization ensures that the reconstruction results conform to the physical characteristics of the transmission line by introducing PnP constraints and spatial straight line constraints. Finally, dynamic update processing realizes the temporal continuity of depth estimation by considering historical depth information, and outputs depth uncertainty and consistency indicators at the same time, providing reliable confidence quantification for subsequent state evaluation.

[0099] According to one aspect of the present application, step S32 is further:

[0100] S321, based on the depth probability data, using the AKAZE algorithm to extract scale-invariant features to obtain local feature data; performing non-maximum suppression on the local feature data to generate key point data; based on the key point data and confidence data, performing feature point screening and outputting feature point data;

[0101] S322, based on the feature point data, construct a local descriptor based on hash coding to obtain descriptor data; based on the descriptor data, use local sensitive hashing to perform fast neighbor search to generate candidate matching data; apply geometric consistency constraints to the candidate matching data, and output feature matching data;

[0102] S323, based on the feature matching data and the confidence data, calculate the local distribution characteristics of the matching points to obtain first distribution feature data; perform reliability assessment on the first distribution feature data based on kernel density estimation to generate reliability data; perform weighted fusion based on the reliability data and the confidence data to output feature weight data;

[0103] S324. Based on the feature matching data and feature weight data, a global optimization model based on graph cuts is constructed to obtain optimization model data; the optimization model data is iteratively solved to generate iterative optimization data; the iterative optimization data is refined in combination with spatial consistency constraints to output feature depth data.

[0104] In one embodiment of the present application, the AKAZE feature extraction improved algorithm: R(x, y, σ) = det(H) - k·tr(H) 2·exp(-α·L(x,y))·(1 + β·S(σ)); where H is the Hessian matrix; det(H) is the determinant; tr(H) is the trace; L(x,y) = ||▽I|| 2 is the gradient amplitude; S(σ) = σ / σ_max is the scale ratio; k is the weight parameter; α is the gradient weight coefficient, ranging from [0.1, 2]; β is the scale gain coefficient, ranging from [0, 1]; σ_max is the maximum scale; R(x, y, σ) is the response function of the coordinate point (x, y) at the scale σ; ▽I is the image gradient.

[0105] Hash coded local descriptor: H(p) = sign(∑w_i·f_i(p))·exp(-α·D(p))·(1 +β·R(p)); where f_i(p) is the local feature at feature point p; w_i is the feature weight; D(p) = ||p - p_c|| 2 is the spatial distance term; R(p) = ||▽I(p)|| 2 is the gradient response; p_c is the feature center point; sign(·) is the sign function; α is the distance attenuation coefficient, ranging from [0.1, 1]; β is the gradient gain coefficient, ranging from [0, 2]; H(p) is the hash code at the feature point p; I(p) is the image intensity at the feature point p.

[0106] Graph cut global optimization model: E(f) = ∑D_p(f_p) + λ·∑V_pq(f_p,f_q)·exp(-α·B(p,q))·(1 + β·S(p,q)); where D_p(f_p) is the data term; V_pq(f_p,f_q) is the smoothing term; B(p,q) = ||I(p)-I(q)|| 2 is the brightness difference; S(p,q) = cos(θ_p,θ_q) is the direction similarity; f_p, f_q are labels; θ_p, θ_q are direction vectors; λ is the smoothing coefficient; α is the brightness weight coefficient, ranging from [0.1, 2]; β is the direction weight coefficient, ranging from [0, 1]; E(f) is the global energy function; I(p) and I(q) are the image intensities at points p and q, respectively.

[0107] This embodiment achieves high-precision depth information extraction through a multi-level feature matching strategy. The scale-invariant features provided by the AKAZE algorithm are highly distinguishable and are particularly suitable for feature matching under perspective transformation. The hash-coded local descriptor reduces the computational complexity of feature storage and matching while maintaining good matching performance. The application of local sensitive hashing improves the efficiency of neighbor search, especially when processing large-scale feature points. The introduction of geometric consistency constraints effectively filters out false matches and improves the reliability of matching. The kernel density estimation method provides a theoretical basis for feature point reliability assessment, while the graph cut global optimization ensures the overall consistency of feature matching results. The application of spatial consistency constraints further improves the accuracy of depth estimation. This embodiment not only ensures the accuracy of feature matching, but also achieves efficient computational processing.

[0108] According to one aspect of the present application, step S33 is further:

[0109] S331, read feature depth data, use wavelet transform to perform multi-resolution decomposition to obtain wavelet coefficient data; perform adaptive threshold filtering on the wavelet coefficient data to generate filter coefficient data; reconstruct multi-scale representation based on the filter coefficient data, and output depth pyramid data;

[0110] S332, obtaining depth pyramid data, constructing an inter-layer relationship model based on conditional random fields to obtain inter-layer model data; calculating a potential energy function for the inter-layer model data to generate potential energy distribution data; performing probabilistic inference on the potential energy distribution data in combination with spatial consistency constraints, and outputting depth propagation data;

[0111] S333, receiving the depth propagation data, constructing the objective function based on the variational method to obtain the variational function data; performing L0 gradient minimization on the variational function data to generate gradient feature data; analyzing the anisotropy of the gradient feature data in combination with the structure tensor, and outputting the optimized target data;

[0112] S334, read the optimization target data and the depth propagation data, use the alternating direction multiplier method to optimize and solve to obtain the iterative result data; perform scale space consistency verification on the iterative result data to generate consistency index data; optimize the depth value based on the consistency index data, and output consistent depth data.

[0113] This embodiment achieves consistency optimization of depth information through multi-scale decomposition and probabilistic inference. The multi-resolution decomposition of wavelet transform provides a hierarchical representation of depth information, and the adaptive threshold filtering effectively suppresses noise at each scale. The conditional random field model describes the inter-layer relationship of depth information, making the depth propagation more in line with physical laws. The design of the potential energy function takes into account the spatial continuity constraint to ensure the smooth transition of depth information. The construction of the variational objective function combines the data term and the regularization term, and the L0 gradient minimization effectively maintains the step characteristics of the depth. Structural tensor analysis provides directional information of depth changes, and the alternating direction multiplier rule realizes efficient optimization solution. Scale space consistency verification ensures the consistency of depth information at different scales. This embodiment not only ensures the accuracy of depth estimation, but also realizes the effective fusion of multi-scale information.

[0114] like Figure 5 As shown, according to one aspect of the present application, step S4 is further:

[0115] S41. Based on the optimized depth data and the pre-stored historical environmental parameters, a generalized catenary model is constructed to obtain catenary model data; based on the catenary model data, environmental impact factors are calculated and environmental impact data are generated; based on the historical environmental parameters and environmental impact data, an adaptive relationship is constructed to output physical constraint data;

[0116] S42, based on the physical constraint data and the pre-stored physical model parameters, calculate the stress-strain relationship of the conductor to obtain strain distribution data; based on the feature aggregation data and the physical model parameters, calculate the curvature of the conductor; based on the strain distribution data and the curvature of the conductor, calculate the tension distribution and generate tension distribution data; based on the physical constraint data, construct an energy constraint term; combine the tension distribution data with the energy constraint term and output dynamic tension data;

[0117] S43, based on the optimized depth data and the depth consistency data, construct a spatiotemporal covariance matrix to generate covariance data; based on the covariance data, calculate the continuity score of the spatial position to obtain continuity data; perform smoothing constraint processing on the covariance data and the continuity data, and output spatiotemporal constraint data;

[0118] S44. Based on the dynamic tension data, spatiotemporal constraint data and depth uncertainty data, a second constraint energy function is constructed to generate second constraint energy data; based on the second constraint energy data, dynamic weights are calculated to obtain weight update data; based on the weight update data, the second constraint energy data is iteratively optimized, and finally system state data for system evaluation, three-dimensional point cloud data for three-dimensional reconstruction and deformation parameter data for deformation analysis are output.

[0119] In one embodiment of the present application, the physical constraint energy functional algorithm: E_total = ∫(T·κ(s) 2 +EI·τ(s) 2 + ρ·g·h(s))ds + α·∫(ΞT / Ξs) 2 ds + β·∫∫σ(x,y)·ε(x,y)dxdy; where T is tension; κ(s) is curvature; τ(s) is torsion; EI is bending stiffness; ρ is linear density; g is gravitational acceleration; h(s) is height function; σ(x,y) is stress; ε(x,y) is strain; α is the tension change penalty coefficient, ranging from [0.1, 1]; β is the stress-strain coupling coefficient, ranging from [0.01, 0.5]; E_total is the total energy functional; s is the curve parameter.

[0120] Spatiotemporal covariance modeling algorithm: K(x, t; x', t') = k_s(x, x')·k_t(t, t')·exp(-α·D(x, x', t, t'))·(1 + β·C(x, x', t, t')); where k_s(x, x') = exp(-||x-x'|| 2 / 2l_s 2 ) is the spatial kernel function; k_t(t, t') = exp(-|t-t'| 2 / 2l_t 2 ) is the time kernel function; D(x, x', t, t') = ||v(x, t)-v(x', t')|| 2 is the velocity difference term; C(x, x', t, t') = cos(θ(x, t), θ(x', t')) is the directional consistency term; v(x, t) is the velocity vector; θ(x, t) is the directional angle; l_s is the spatial scale parameter; l_t is the time scale parameter; α is the velocity weight coefficient, ranging from [0.1, 2]; β is the direction weight coefficient, ranging from [0, 1]; K(x, t; x', t') is the space-time covariance function; x, x' are spatial coordinates; t, t' are time coordinates.

[0121] This embodiment combines the physical model with the data-driven method to achieve a comprehensive analysis and evaluation of the state of the transmission line system. Through the construction of the physical model, the mechanical behavior of the conductor under different environmental conditions is accurately simulated, especially considering the influence of environmental factors such as temperature and wind load on the conductor morphology. The tension deformation analysis is based on the theory of nonlinear mechanics and accurately calculates the stress-strain state of each point on the conductor, providing a reliable basis for evaluating the safety of the conductor. The spatiotemporal consistency processing effectively suppresses the influence of measurement noise and environmental disturbances by establishing a spatiotemporal correlation model of the conductor state. The physical constraint fusion optimization combines data-driven observation results with physical model predictions, and achieves high-precision state estimation through iterative optimization. The final output system state data, three-dimensional point cloud data, and deformation parameter data fully describe the working state of the conductor system, providing reliable data support for operation and maintenance decisions.

[0122] According to one aspect of the present application, step S42 is further:

[0123] S421, reading physical constraint data and physical model parameters, constructing a wire mechanical model using a finite element method to obtain mechanical model data; performing nonlinear stress analysis on the mechanical model data to generate stress field data; calculating strain distribution based on a constitutive equation, and outputting strain distribution data;

[0124] S422, obtaining strain distribution data, using a B-spline curve to fit the shape of the wire to obtain curve fitting data; calculating the local curvature of the curve fitting data to generate curvature distribution data; calculating the tension value of each point in combination with material parameters, and outputting the tension distribution data;

[0125] S423, receiving tension distribution data, constructing an energy functional based on Hamiltonian principle to obtain energy functional data; performing variational solution on the energy functional data to generate variational result data; considering boundary condition constraints, outputting constraint result data;

[0126] S424, reading the constraint result data and the tension distribution data, and performing iterative solution using the dynamic relaxation method to obtain iterative solution data; performing error analysis on the iterative solution data to generate error evaluation data; performing tension correction based on the error evaluation data, and outputting dynamic tension data.

[0127] This embodiment achieves an accurate assessment of the stress state of the wire through sophisticated mechanical modeling and dynamic tension analysis. The wire mechanical model constructed by the finite element method can accurately describe the nonlinear mechanical behavior of the wire, especially when considering large deformation and geometric nonlinearity. The application of stress field analysis and constitutive equations makes the strain distribution calculation more consistent with the actual physical process. B-spline curve fitting not only provides a continuous expression of the wire shape, but also provides a geometric basis for tension estimation through curvature calculation. The energy functional construction based on Hamilton's principle introduces the system dynamics into the tension analysis, making the calculation results more consistent with the principle of energy conservation. The variational solution process takes into account the boundary condition constraints to ensure the physical rationality of the results. The application of dynamic relaxation method improves the stability of iterative solution, while error analysis and tension correction further improve the reliability of the results. This embodiment not only ensures the accuracy of mechanical analysis, but also realizes the numerical stability of the calculation process.

[0128] According to one aspect of the present application, step S43 is further:

[0129] S431, read the optimized depth data and the depth consistency data, construct a time series sliding window to obtain time series window data; perform kernel density estimation on the time series window data to generate density distribution data; calculate the covariance matrix of the density distribution data based on the Mahalanobis distance, and output the covariance data;

[0130] S432, obtaining optimized depth data, using a time series differential operator to calculate the displacement field to obtain displacement field data; performing curvature analysis on the displacement field data to generate curvature feature data; scoring the curvature feature data in combination with local consistency constraints, and outputting continuity data;

[0131] S433, receiving covariance data, constructing a covariance estimation model based on Wishart distribution to obtain estimation model data; performing sparsity optimization on the estimation model data to generate sparse structure data; regularizing the sparse structure data based on a graph Laplace operator, and outputting smooth matrix data;

[0132] S434, read smooth matrix data and continuity data, use spatiotemporal tensor decomposition to obtain tensor decomposition data; perform rank constraint optimization on the tensor decomposition data to generate optimized tensor data; combine multi-scale consistency constraints to fuse the optimized tensor data, and output spatiotemporal constraint data.

[0133] This embodiment achieves an accurate description of the motion state of the wire by constructing a spatiotemporal covariance structure and continuity constraints. The design of the time-series sliding window enables the system to effectively capture the time-series characteristics of the wire motion, while the kernel density estimation provides a non-parametric representation of the data distribution. The introduction of the Mahalanobis distance makes the covariance calculation more robust to outliers. The displacement field calculated by the time-series differential operator directly reflects the motion characteristics of the wire, and the curvature analysis provides a geometric measure of the motion continuity. The application of the Wishart distribution provides a theoretical basis for covariance estimation, and the sparsity optimization effectively reduces the complexity of the model. The regularization of the graph Laplacian operator ensures smoothness in space, and the spatiotemporal tensor decomposition achieves unified processing in time and space dimensions. The introduction of rank-constrained optimization effectively controls the complexity of the model, and the multi-scale consistency constraint ensures the consistency of state estimation at different scales. This embodiment not only ensures the accuracy of state estimation, but also realizes the effective fusion of spatiotemporal features.

[0134] According to one aspect of the present application, the method further includes: S5, performing multi-dimensional state evaluation based on system state data and deformation parameter data to generate state evaluation data; performing anomaly detection analysis based on the state evaluation data and three-dimensional point cloud data to obtain anomaly detection data; performing early warning decision processing based on the anomaly detection data, state evaluation data and pre-stored early warning threshold parameters, and outputting state report data, early warning information data and decision suggestion data;

[0135] S51, based on the system state data and deformation parameter data, extracting geometric features and physical features to obtain feature index data; based on the feature index data, calculating risk assessment indexes to generate risk index data; performing weighted combination of the feature index data and the risk index data to output state assessment data;

[0136] S52, based on the state assessment data and the three-dimensional point cloud data, calculating data distribution characteristics to obtain second distribution characteristic data; based on the second distribution characteristic data, identifying abnormal patterns and generating abnormal probability data; based on the abnormal probability data, classifying and locating the identified abnormalities and outputting abnormality detection data;

[0137] S53. Based on the anomaly detection data, the status assessment data and the pre-stored warning threshold parameters, the warning level is calculated to obtain the warning level data; the current system context is obtained, and based on the warning level data and the current system context, a decision suggestion is generated to obtain the decision generation data; based on the anomaly detection data, the status assessment data, the pre-stored warning threshold parameters, the warning level data and the decision generation data, a system status report is formed, and finally the status report data for the operation and maintenance personnel, the warning information data for the warning system and the decision recommendation data for the decision support are output.

[0138] In one embodiment of the present application, a multi-feature fusion state evaluation algorithm is: S(x) = ω_g·G(x)·exp(-α·E_g(x)) + ω_p·P(x)·exp(-β·E_p(x)) + ω_d·D(x)·exp(-γ·E_d(x)); wherein G(x) is a geometric feature score; P(x) is a physical feature score; D(x) is a dynamic feature score; E_g(x) is a geometric feature error; E_p(x) is a physical feature error; E_d(x) is a dynamic feature error; ω_g, ω_p, ω_d are feature weights, satisfying ∑ω_i=1; α is a geometric error coefficient, ranging from [0.5, 2]; β is a physical error coefficient, ranging from [0.1, 1]; γ is a dynamic error coefficient, ranging from [0.2, 1.5]; S(x) is a comprehensive state evaluation score; x is a state vector.

[0139] Anomaly detection probability model algorithm: P(a|x) = π_0·N(x|μ_0,Σ_0)·exp(-α·R(x)) + π_1·M(x|θ)·exp(-β·D(x)); where π_0 and π_1 are mixing ratios; N(x|μ_0,Σ_0) is the normal state Gaussian distribution; M(x|θ) is the abnormal state distribution; R(x) is the residual term; D(x) is the distance metric; α is the residual weight coefficient, ranging from [0.1, 2]; β is the distance weight coefficient, ranging from [0.5, 5]; θ is the abnormal distribution parameter vector; P(a|x) is the posterior probability of the abnormal state a given the observation x; μ_0 is the normal state mean vector; Σ_0 is the normal state covariance matrix.

[0140] This embodiment ensures a comprehensive assessment of the system status by generating comprehensive status assessment data, and can timely discover potential risks and anomalies; by outputting anomaly detection data, the accuracy and reliability of anomaly detection are improved, and anomalies in the system can be effectively identified and located; by integrating all information to form system status reports, warning information and decision-making recommendation data, the timeliness and accuracy of warning decisions are ensured, and effective support can be provided for operation and maintenance personnel and warning systems; through multi-dimensional status assessment and anomaly detection analysis, efficient monitoring and management capabilities can be maintained in complex and dynamic environments, ensuring the robustness and reliability of decisions. The system status report data, warning information data and decision-making recommendation data finally output provide detailed system status information and warning prompts for operation and maintenance personnel, and provide a scientific decision-making basis for the decision support system. This embodiment realizes comprehensive monitoring and management of the system status through systematic evaluation, detection and decision-making processing, and improves the security and reliability of the system.

[0141] According to one aspect of the present application, step S51 is further:

[0142] S511, receiving system state data and deformation parameter data, extracting geometric invariants using principal curvature analysis to obtain geometric feature data; performing stress-strain analysis on the geometric feature data to generate physical feature data; performing dimensionality reduction optimization on the geometric feature data and the physical feature data based on feature correlation analysis, and outputting feature index data;

[0143] S512, acquiring characteristic indicator data, constructing a risk assessment model based on extreme value statistics to obtain extreme value model data; performing generalized Pareto distribution fitting on the extreme value model data to generate distribution parameter data; combining failure probability calculation and distribution parameter data to perform risk quantification, and outputting risk indicator data;

[0144] S513, read the characteristic indicator data and risk indicator data, use the fuzzy analytic hierarchy process to construct a weight system to obtain weight hierarchy data; perform consistency check on the weight hierarchy data to generate consistency indicator data; optimize weight allocation based on the consistency indicator data, and output combined weight data;

[0145] S514, receiving the combined weight data, characteristic index data and risk index data, constructing a multi-criteria fusion model based on Choquet integral to obtain fusion model data; performing interactive characteristic analysis on the fusion model data to generate interactive evaluation data; performing comprehensive evaluation on the interactive evaluation data in combination with fuzzy measurement theory, and outputting state evaluation data.

[0146] In one embodiment of the present application, the principal curvature calculation model K(p) = (λ_1·k_1(p) + λ_2·k_2(p))·exp(-α·H(p))·(1 + β·G(p)); where k_1(p) and k_2(p) are the principal curvature values ​​at point p, k_i =(-b_i ± sqrt(b_i 2 - 4a_i·c_i)) / (2a_i); H(p) = (k_1(p) + k_2(p)) / 2 is the mean curvature; G(p) = k_1(p)·k_2(p) is the Gaussian curvature; a_i, b_i, c_i are the quadratic surface fitting coefficients; λ_1 and λ_2 are the principal curvature weight coefficients, satisfying λ_1 + λ_2 = 1; α is the mean curvature influence coefficient, ranging from [0.1, 1]; β is the Gaussian curvature gain coefficient, ranging from [0, 2].

[0147] Extreme value statistical risk assessment model: R(x) = (1-exp(-((x-μ) / σ) k))·exp(-α·T(x))·(1 +β·V(x)); where x is the observation value; μ is the location parameter; σ is the scale parameter; k is the shape parameter; T(x) = (x-x_th) / x_th is the threshold exceedance rate; V(x) = var(x) / var_max is the normalized variance; x_th is the threshold; α is the exceedance rate weight coefficient, ranging from [0.5, 2]; β is the variance gain coefficient, ranging from [0, 1]; R(x) is the risk assessment function; var_max is the maximum variance.

[0148] Fuzzy hierarchical analysis weight optimization: w_i = (∏a_ij 1 / n ) / (∑(∏a_ij 1 / n ))·exp(-α·C(i))·(1 + β·E(i)); where a_ij is the element of the fuzzy judgment matrix; n is the number of criteria; C(i) = |CI_i| / RI is the consistency ratio; E(i) = e_i / e_max is the expert score; CI_i is the consistency index; RI is the random consistency index; α is the consistency weight coefficient, ranging from [0.1, 1]; β is the expert weight coefficient, ranging from [0, 2]; w_i is the weight of the i-th criterion; e_max is the maximum expert score, ∏ is the product operator, and e_i is the score of the i-th expert.

[0149] Choquet integral multi-criteria fusion: C_μ(f) = ∑(f(x_i)-f(x_(i-1)))·μ(A_i)·exp(-α·D(A_i))·(1 + β·I(A_i)); where f(x_i) is the scoring function; μ(A_i) is the fuzzy measure; A_i is the indicator set; D(A_i) = |μ(A_i) - μ_ref| is the measurement deviation; I(A_i) = ∑(μ(B∪{i}) - μ(B)) is the interaction index; α is the deviation penalty coefficient, ranging from [0.1, 2]; β is the interaction gain coefficient, ranging from [0, 1]; C_μ(f) is the Choquet integral value; μ_ref is the reference fuzzy measure value; B is the indicator subset.

[0150] This embodiment achieves a comprehensive evaluation of the system state through multi-dimensional feature extraction and comprehensive evaluation. The geometric invariants extracted by the principal curvature analysis provide a stable feature representation for state evaluation, and the stress-strain analysis reflects the mechanical state of the wire. The dimensionality reduction optimization of the feature correlation analysis effectively reduces redundant features and improves the efficiency of the evaluation. The extreme value statistical model analyzes the limit state of the system, and the generalized Pareto distribution fitting provides a reliable probability model for risk assessment. The failure probability calculation takes into account a variety of failure modes, making the risk quantification more comprehensive. The weight system constructed by the fuzzy hierarchical analysis method fully considers the relative importance of different indicators, and the consistency test ensures the rationality of the weight distribution. The multi-criteria fusion based on Choquet integral considers the interaction between the evaluation indicators, and the application of fuzzy measurement theory makes the comprehensive evaluation results more objective and reliable. This embodiment not only ensures the comprehensiveness of the state evaluation, but also realizes the reliability of the evaluation results.

[0151] According to one aspect of the present application, step S52 is further:

[0152] S521, reading the state assessment data and the three-dimensional point cloud data, using local feature statistics to calculate statistical feature data; performing multi-dimensional noise reduction processing on the statistical feature data to generate noise reduction feature data; using kernel principal component analysis to extract the main distribution features, and outputting the distribution feature data;

[0153] S522, obtaining distribution feature data, constructing an anomaly detection model based on an isolation forest to obtain detection model data; performing adaptive threshold analysis on the detection model data to generate threshold parameter data; calculating anomaly probability by combining a density estimation method, and outputting anomaly probability data;

[0154] S523, receiving abnormal probability data, using a spectral clustering algorithm to cluster abnormal patterns to obtain clustering result data; performing spatiotemporal correlation analysis on the clustering result data to generate correlation data; constructing an abnormal propagation graph based on the correlation data, and outputting the propagation graph data;

[0155] S524. Read the propagation graph data and anomaly probability data, use the graph neural network to perform anomaly classification to obtain classification result data; perform spatial positioning calculation on the classification result data to generate positioning data; combine the positioning data and anomaly type information to output anomaly detection data.

[0156] This embodiment achieves accurate detection and positioning of system anomalies through deep learning and statistical analysis methods. The calculation of local feature statistics provides statistical characteristics of state data, and multi-dimensional noise reduction processing effectively suppresses data noise. The application of kernel principal component analysis not only realizes nonlinear feature extraction, but also maintains the essential characteristics of the data. The construction of the isolation forest model utilizes the isolation characteristics of the data for anomaly detection, and the adaptive threshold analysis improves the accuracy of detection. The density estimation method provides a theoretical basis for the calculation of anomaly probability. The application of spectral clustering algorithm realizes the automatic classification of abnormal patterns, and the spatiotemporal correlation analysis reveals the law of abnormal propagation. The introduction of graph neural network enables the system to make full use of the correlation between anomalies for classification, and the spatial positioning calculation realizes the accurate positioning of anomalies. This embodiment not only ensures the accuracy of anomaly detection, but also realizes the effective analysis of the law of abnormal propagation.

[0157] In one embodiment of the present application, a method for three-dimensional reconstruction of a power transmission line based on a monocular channel monitoring camera includes the following steps:

[0158] Step 1: Based on the geometric constraints of the wire in the image, the three-dimensional contour of the wire is preliminarily constructed to ensure the accuracy and consistency of the three-dimensional reconstruction;

[0159] If there are multiple pairs of wires (more than two) in the image captured by the camera, and the expressions of these wires in the image are known, i x+b i y+c i = 0, the camera internal parameter K and the relative position relationship between the ith wire and the i+1th wire in space (expressed by a three-dimensional vector) vec i , then these wires will satisfy certain geometric constraints, and the wires can be directly reconstructed in three dimensions through these geometric constraints. Taking three wires as an example, assuming that the three wires are parallel to each other in space, a vanishing point is derived from the three wires. Let op ij =l i × j , where × represents the vector outer product operation, l i =(a i , b i , c i ), is a three-dimensional vector consisting of the wire expression parameters. Then, vp x =1 / N∑(op ij1 / op ij3 ), vp y =1 / N∑(op ij2 / op ij3 ), where N is the number of wire pairs, and in this embodiment, N=3.

[0160] From the definition of the vanishing point, we can know that the line l connecting the vanishing point vp and the origin O of the camera coordinate system is o-vp The direction is the wire direction in the camera coordinate system, and l o-vp =λ[(vp x -c x ) / f x ,(vp y -c y ) / f y , 1] T , where λ∈R, f x 、f y 、c x and c y is the camera internal parameter.

[0161] The origin and the three planes P1, P2, and P3 drawn from the three end points of the three wires on the imaging plane intersect at l o-vp . At a point perpendicular to l o-vp On the plane, l o-vp The three wires form four vertical points pl, p1, p2, and p3 on the plane. p1, p2, and p3 intersect with the plane to form three straight lines; from the geometric relationship, we know that p1, p2, and p3 can only be on the three black dotted lines, and the vertical points must satisfy the known relative position relationship. The geometric relationship between pl and p1, p2, and p3 is deduced below. α1, α2 are straight lines l pl-p1 , l pl-p2 , l pl-p3 Let the direction vector of wire i in the image be vec wi , straight line l o-vp The direction vector is vec vp , then the straight line l pl-pi The direction vector vec li =vec wi -(vec wi ·vec vp / |vec vp ∣) vec vp , thus we get: αi=arccos(vec li ·vec li+1 / |vec li ∣·∣vec li+1∣); β1 is the complementary angle of the angle between vec1 and vec2, that is, β1= arccos(vec1·vec2 / ∣vec1∣·∣vec2∣). l1=∣vec1∣, l2=∣vec2∣, and θ is the variable to be determined. Let the distance between p1 and p2 be x, then, from the law of sines, we can get: l1 / sin(α1)=x / sin(π-θ1); l2 / sin(α2)=x / sin(π+θ1-φ); where φ=α1+α2+β1; Combining the above two equations and eliminating x, we can get: θ=arctan(sin(φ) / (r+cos(φ))), where r= l1 sin(α2) / l2 sin(α1); Since θ needs to satisfy 0≤θ≤ π, θ has a unique solution. We can then find p l The distances d1, d2, and d3 between p1, p2, and p3 are: d1 = l1 sin(θ1-α1) / sin(α1); d2 = l1 sin(π-θ1) / sin(α1); d3 = l12sin(π-θ2) / sin(α2); where θ2 = π+θ1-(α1+β1). When there are more than three wires, the wires can be split into groups of three, and the above conditions can still be met. At this time, N w -2 θ values, where N w is the number of wires. And because the values ​​of θ satisfy θi+1=π+θi-(αi+βi). Based on this relationship, N can be recursively obtained. w -2 θ1 values, the final θ1=1 / (N w -2) ∑θ li Similarly, based on this relationship, the final θi value can be recursively obtained, and then p can be obtained. l The distance di between the i-th wire and pi. Finally, the three-dimensional expression of the i-th wire in the camera coordinate system is di*(vec li / |vec li ∣)+ l o-vp .

[0162] Step 2: By using the PnP algorithm, the depth information of each point on the wire is optimized by combining the known locations of the wire feature points with the camera internal parameters;

[0163] For a pair of wires, the PnP problem is introduced to further optimize its depth information: Taking the wire pair consisting of the first wire and the second wire as an example, let a pixel point on the first wire in the image be p1, and calculate the coordinates of the wire point in the three-dimensional space corresponding to the point in the camera coordinate system through the depth dep1 corresponding to the point and the camera intrinsic parameter K P1 = dep1*K T *[p 1x , p 1y , 1] TIn step 1, the expression of the i-th wire in the camera coordinate system is known, so the expression of the vector vec1 representing the relative position relationship between the first wire and the second wire in the camera coordinate system is: vec 1c =d2* vec l2 / |vec l2 ∣- d1* vec l1 / |vec l1 ∣=[d x1 , d y1 , d z1 ] T ; So the point P2 on the second wire that matches P1 in the camera coordinate system = P1 + vec 1c .

[0164] The vector vec is calculated by the definition of the vanishing point. 1c The vanishing point formed by the parallel straight lines in the camera image is vp vec1 (f x (d x1 / d z1 )+c x , f y (d y1 / d z1 )+c y ). The expression of the second wire on the image is known to be a2x+b2y+c2=0, let op=[p 1x , p 1y , 1] T ×[vp vec1，x , vp vec1，y , 1] T ×[a2,b2,c2] T Then, the matching point of pixel p1 on the second wire in the image is p2=(op1 / op3, op2 / op3). Therefore, according to the PnP problem, the optimized depth corresponding to all pixels on the first wire can be expressed as: dep1=argmin dep1 1 / 2∑ i=1 N ∣∣p 2i -1 / s i K(dep 1i *K T *[p 1ix , p 1iy , 1] T +vec 1c ) ∣∣2 2 ; Where N is the number of pixels on the wire, s i For point p 2i The z coordinate p 2iz =(dep1i * K T *[p 1ix , p 1iy , 1] T +vec 1c ).

[0165] Step 3: Considering the natural catenary shape of the wire under the action of gravity, the catenary model is used to optimize the wire depth information, especially to provide more accurate estimation on the depth information of long-span wires, and enhance the reliability and practicality of 3D reconstruction. Specifically:

[0166] Considering that the transmission line with both ends tied to the transmission tower is a catenary in its natural state, a catenary model is introduced to optimize the depth information of a certain line. Since the introduction of the catenary model requires that the line connecting the two end points of the line is parallel to the ground, and currently only the expression of the line in the camera coordinate system is known, the above premise cannot be guaranteed. The line needs to be transformed from the camera coordinate system to the geodetic coordinate system with the ground as the reference system. When the camera takes a bird's-eye view of the transmission line, the image contains the ground part. The ground depth dep is obtained using the monocular depth estimation and ground segmentation model. g The pixel position p corresponding to the ground g , and then calculate the ground point cloud P g , where: P gi =dep g * K -1 *[p gx , p gy , 1] T , P gi ∈P g .

[0167] Using the ground point cloud, the plane expression of the ground in the camera coordinate system is calculated according to the random sampling consensus (RANSAC) method, and then the expression of the ground normal vector in the camera coordinate system vec is obtained gc . Let the ground normal vector in the geodetic coordinate system be vec g =[0, -1, 0] T , then the coordinate transformation from the camera coordinate system to the earth coordinate system can be expressed as a rotation vector vec rot =vec gc ×vec g .

[0168] For a pixel on the wire in the image, the corresponding perspective transformation is M=KRK -1 , where the rotation matrix R from the camera coordinate system to the earth coordinate system can be obtained by the corresponding rotation vector vec according to the Rodriguez formula rot Calculated.

[0169] The pixel point p on the wire after perspective transformation in the image i And the corresponding depth dep i , calculate the three-dimensional coordinates of the point on the wire in the geodetic coordinate system: P i =dep i * K -1 *[p ix , p iy , 1] T =[x i ,y i , z i ] T Then, according to the catenary equation, the catenary coefficient of a certain point on the wire is calculated as η i =y i n / sqrt(1+y' i 2 ), where y' i = (y' i+1 - y' i ) / (d xzi+1 -d xzi ), d xzi+1 =sqrt(x i 2 +z i 2 ). The optimized depth corresponding to all pixels is obtained by making the catenary coefficients at all locations on the catenary equal: dep = argmin dep 1 / 2∑ i=1 N-3 (η i+1 -η i ) 2 It should be pointed out that the optimization schemes proposed in step 2 and step 3 can also be used as constraints to optimize the wire depth.

[0170] According to one aspect of the present application, a three-dimensional reconstruction system of a power transmission line based on a monocular channel monitoring camera includes:

[0171] at least one processor; and,

[0172] a memory communicatively connected to at least one of the processors; wherein,

[0173] The memory stores instructions that can be executed by the processor, and the instructions are used to be executed by the processor to implement the three-dimensional reconstruction method of the transmission line based on the monocular channel monitoring camera described in any of the above embodiments.

[0174] The present invention realizes high-precision three-dimensional reconstruction and status monitoring of power transmission lines based on a monocular camera by constructing a complete processing flow from image acquisition to status assessment. The multi-level image processing and feature extraction strategy effectively overcomes the imaging interference in complex environments; the spatial structure reconstruction method based on topological and geometric dual constraints ensures the accuracy and physical rationality of the reconstruction results; the depth estimation and uncertainty quantification method under the probabilistic framework provides reliable depth information and confidence assessment; the physical model-driven status assessment method realizes accurate analysis of the status of the conductor system. The present invention has strong engineering practicality, can adapt to complex field environments, and provide reliable monitoring results. In particular, it shows obvious advantages when dealing with challenging scenarios such as large spans, multiple conductors, and complex meteorological conditions. The present invention provides strong technical support for the intelligent operation and maintenance of transmission lines.

[0175] The preferred embodiments of the present invention are described in detail above; however, the present invention is not limited to the specific details in the above embodiments. Within the technical concept of the present invention, various equivalent transformations can be made to the technical solutions of the present invention, and these equivalent transformations all belong to the protection scope of the present invention.

Claims

1. A three-dimensional reconstruction method for power transmission lines based on a monocular channel monitoring camera, characterized in that: The steps include: S1, obtaining original camera image data, generating compensated image data through adaptive illumination compensation processing; performing frequency domain feature analysis on the compensated image data to obtain directional feature data; Based on the directional feature data, a graph theory method is used to build a texture enhancement model to generate enhanced image data; based on the enhanced image data, time series feature aggregation processing is performed to output feature aggregation data; S2, based on the feature aggregation data and the directional feature data, generate topology constraint data through multi-conductor topology analysis; based on the directional feature data and the feature aggregation data, extract geometric constraint conditions; Based on the topological constraint data and the geometric constraint conditions, the first constraint energy data is constructed, and deformation compensation processing is performed to obtain deformation compensation data; based on the deformation compensation data, optimized constraint data is generated through constraint optimization calculation; based on the optimized constraint data, robustness verification is performed to output three-dimensional contour data; S3, based on the three-dimensional contour data, generate depth probability data through depth probability modeling; based on the depth probability data, perform feature matching optimization to obtain feature depth data; Based on the feature depth data, multi-scale consistency processing is performed to generate consistent depth data; based on the consistent depth data and pre-stored camera calibration parameters, geometric constraint optimization is performed to obtain constrained depth data; based on the constrained depth data, optimized depth data, depth uncertainty data and depth consistency data are output through dynamic update processing; S4. Based on the optimized depth data and pre-stored historical environment parameters, physical constraint data is constructed through a pre-configured physical model; based on the physical constraint data and pre-stored physical model parameters, dynamic tension data is obtained through tension deformation analysis; based on the optimized depth data and depth consistency data, spatiotemporal constraint data is generated through spatiotemporal consistency processing; Based on dynamic tension data, spatiotemporal constraint data and depth uncertainty data, through physical constraint fusion optimization, the system state data, three-dimensional point cloud data and deformation parameter data are finally output; Step S1 is further as follows: S11, obtaining original camera image data, calculating image brightness distribution, and obtaining brightness histogram data; performing local area entropy analysis on the brightness histogram data to generate local entropy weight data; calculating a compensation coefficient based on the local entropy weight data; performing pixel-level multiplication operation on the compensation coefficient and the original camera image data, and outputting compensated image data; S12, performing a two-dimensional Fourier transform based on the compensated image data to obtain frequency domain feature data; Constructing a predetermined number of directional filters to generate a directional filter group; performing a convolution operation on the frequency domain feature data and the directional filter group to obtain directional response data; performing maximum value extraction on the directional response data to output directional feature data; S13, constructing a pixel association graph based on the compensated image data and the directional feature data, and generating graph structure data; Calculate the edge weights in the graph structure data to obtain the weight matrix data; Perform eigenvalue decomposition on the weight matrix data to obtain eigenvector data; Based on the feature vector data, the compensated image data is reconstructed to output enhanced image data; S14, based on the enhanced image data, obtaining an enhanced image data sequence within a preset time window, and constructing a time series feature tensor; Calculate the feature similarity of adjacent moments in the time series feature tensor to generate time series weight data; perform weighted combination of the time series weight data and the time series feature tensor to output feature aggregation data; Step S3 is further as follows: S31, based on the three-dimensional contour data, calculating the initial depth distribution of each point to generate initial depth distribution data; based on the initial depth distribution data, extracting the depth relationship between adjacent points, constructing a depth consistency constraint, and obtaining depth constraint data; combining the initial depth distribution data with the depth constraint data, and outputting depth probability data; S32, extracting feature points based on the depth probability data and the confidence data to obtain feature point data; Calculate the matching relationship of feature point data and generate feature matching data; Based on the feature matching data and confidence data, the reliability weight of the feature points is calculated to obtain feature weight data; Based on the feature matching data and feature weight data, the depth value is optimized and the feature depth data is output; S33, performing multi-scale decomposition on the feature depth data to generate depth pyramid data; Calculate the depth propagation relationship between each level of the depth pyramid data to obtain the depth propagation data; construct an optimization target, optimize the depth propagation data, and output consistent depth data; The optimization objectives include data terms, smoothing terms, and scale terms; S34, constructing PnP constraints based on pre-stored camera calibration parameters and consistent depth data, and generating PnP constraint data; Based on the PnP constraint data, the distance from the spatial point to the straight line is calculated to obtain the geometric consistency score and generate the geometric score data; Fuse the PnP constraint data with the geometric score data, update the depth probability distribution, and output the constrained depth data; S35. Calculate the depth update weight based on the confidence data to obtain update weight data; Based on the updated weight data, the constrained depth data is weightedly combined with the pre-stored historical depth data to generate depth state data; Perform anomaly detection and correction on the depth state data, and finally output optimized depth data, depth uncertainty data, and depth consistency data; Step S11 is further as follows: S111, obtaining original camera image data, and converting it into HSV color space to obtain HSV image data; extracting V channel data in the HSV image data to generate brightness channel data; performing histogram statistical calculation on the brightness channel data, and outputting brightness histogram data; S112, based on the brightness histogram data, using an adaptive window to perform block processing on the image to obtain regional block data; Calculate the grayscale distribution probability of each sub-region in the regional block data to generate probability distribution data; Perform local entropy calculation on probability distribution data and output local entropy weight data; S113, based on the local entropy weight data, construct a weight mapping function based on the local entropy to obtain weight mapping data; perform Gaussian smoothing on the weight mapping data to generate smoothed weight data; based on the smoothed weight data, calculate the compensation coefficient of each pixel position, and output the compensation coefficient data; S114, based on the compensation coefficient data and the original camera image data, respectively mapping the compensation coefficients for the RGB channels to obtain channel compensation data; merging the channel compensation data to generate fused compensation data; Perform edge-preserving filtering on the fused compensation data and output compensated image data.

2. The method for three-dimensional reconstruction of power transmission lines based on a monocular channel monitoring camera according to claim 1, characterized in that: Step S2 is further as follows: S21. Based on the feature aggregation data and the directional feature data, construct a wire spatial parameter equation to obtain spatial parameter data; based on the spatial parameter data, calculate the relative position and angle relationship between the wires to generate position relationship data; extract topological features from the position relationship data, calculate topological invariants, and output topological constraint data; S22, constructing a local coordinate system based on the directional feature data; Projecting the feature aggregation data into a local coordinate system to generate a geometric constraint matrix; calculating a first constraint energy function based on the geometric constraint matrix and the topological constraint data, and outputting first constraint energy data; S23, constructing a wire deformation state space based on the first constraint energy data, and generating state space data; Based on the state space data, the deformation prediction matrix is ​​calculated to obtain the predicted deformation data; Compare the predicted deformation data with the actual observed value, calculate the deformation compensation item, and output the deformation compensation data; S24, constructing a global optimization objective function based on the first constraint energy data, the deformation compensation data and the predefined regularization term to obtain objective function data; Calculate the gradient and quadratic approximation of the objective function data to generate optimization parameter data; use the trust region method to iteratively solve the optimization parameter data and output the optimization constraint data; S25, calculating the structural stability index of the optimized constraint data and generating stability data; Based on the stability data, a confidence map is constructed to obtain confidence data; threshold processing is performed on the confidence data to identify unstable areas; and three-dimensional contour data is generated based on the unstable areas and optimized constraint data.

3. The method for three-dimensional reconstruction of power transmission lines based on a monocular channel monitoring camera according to claim 2, characterized in that: Step S4 is further as follows: S41. Based on the optimized depth data and the pre-stored historical environmental parameters, a generalized catenary model is constructed to obtain catenary model data; based on the catenary model data, environmental impact factors are calculated and environmental impact data are generated; based on the historical environmental parameters and environmental impact data, an adaptive relationship is constructed to output physical constraint data; S42, calculating the stress-strain relationship of the conductor based on the physical constraint data and pre-stored physical model parameters to obtain strain distribution data; Calculate the wire curvature based on feature aggregation data and physical model parameters; Calculate the tension distribution based on the strain distribution data and the curvature of the conductor to generate tension distribution data; Construct energy constraint items based on physical constraint data; Combine the tension distribution data with the energy constraint term to output dynamic tension data; S43, constructing a spatiotemporal covariance matrix based on the optimized depth data and the depth consistency data, and generating covariance data; Based on the covariance data, the continuity score of the spatial position is calculated to obtain the continuity data; the covariance data and the continuity data are smoothed and constrained to output the spatiotemporal constraint data; S44, constructing a second constraint energy function based on the dynamic tension data, the spatiotemporal constraint data and the depth uncertainty data, and generating second constraint energy data; Based on the second constraint energy data, the dynamic weight is calculated to obtain weight update data; Based on the weight update data, the second constraint energy data is iteratively optimized, and finally the system state data for system evaluation, the 3D point cloud data for 3D reconstruction and the deformation parameter data for deformation analysis are output.

4. The method for three-dimensional reconstruction of power transmission lines based on a monocular channel monitoring camera according to claim 3 is characterized in that: Step S23 is further as follows: S231, based on the first constraint energy data, construct a generalized coordinate system through a preconfigured wire mechanics model to obtain coordinate system data; perform modal decomposition on the coordinate system data to generate modal characteristic data; Combine the modal characteristic data with the preconfigured system dynamics equations, construct the state transfer equations, and output the state space data; S232, based on the state space data, using adaptive Kalman filtering to construct a predictor to obtain filter data; performing multi-step prediction on the filter data to generate multi-step prediction data; using Lyapunov stability analysis method to optimize the multi-step prediction data to obtain predicted deformation data; S233, constructing a multi-scale decomposition based on wavelet transform based on the predicted deformation data to obtain scale decomposition data; Perform singular value decomposition on the scale decomposition data to generate singular value data; Based on the singular value data, the main deformation mode is extracted and the deformation mode data is output; S234, based on the deformation pattern data and the predicted deformation data, calculating the residual between the actual observation value and the predicted value to obtain residual vector data; Perform robust M-estimation on the residual vector data to generate robust estimation data; Based on the robust estimation data, an adaptive compensation function is constructed to output deformation compensation data.

5. The method for three-dimensional reconstruction of power transmission lines based on a monocular channel monitoring camera according to claim 3, characterized in that: Step S32 is further as follows: S321, based on the depth probability data, using the AKAZE algorithm to extract scale-invariant features to obtain local feature data; performing non-maximum suppression on the local feature data to generate key point data; based on the key point data and confidence data, performing feature point screening and outputting feature point data; S322, based on the feature point data, construct a local descriptor based on hash coding to obtain descriptor data; based on the descriptor data, use local sensitive hashing to perform fast neighbor search to generate candidate matching data; apply geometric consistency constraints to the candidate matching data, and output feature matching data; S323, based on the feature matching data and the confidence data, calculate the local distribution characteristics of the matching points to obtain first distribution feature data; perform reliability assessment on the first distribution feature data based on kernel density estimation to generate reliability data; perform weighted fusion based on the reliability data and the confidence data to output feature weight data; S324. Based on the feature matching data and feature weight data, a global optimization model based on graph cuts is constructed to obtain optimization model data; the optimization model data is iteratively solved to generate iterative optimization data; the iterative optimization data is refined in combination with spatial consistency constraints to output feature depth data.

6. The method for three-dimensional reconstruction of power transmission lines based on a monocular channel monitoring camera according to claim 3, characterized in that: Also includes: S5. Based on the system state data and deformation parameter data, perform multi-dimensional state assessment to generate state assessment data; Based on the state assessment data and the three-dimensional point cloud data, anomaly detection analysis is performed to obtain anomaly detection data; Based on the abnormal detection data, status assessment data and pre-stored warning threshold parameters, the warning decision processing is carried out to output status report data, warning information data and decision suggestion data; S51, extracting geometric features and physical features based on system state data and deformation parameter data to obtain feature index data; Based on the characteristic indicator data, the risk assessment indicator is calculated to generate risk indicator data; Perform weighted combination of characteristic indicator data and risk indicator data to output status assessment data; S52, based on the state assessment data and the three-dimensional point cloud data, calculating data distribution characteristics to obtain second distribution characteristic data; based on the second distribution characteristic data, identifying abnormal patterns and generating abnormal probability data; Based on the abnormal probability data, the identified anomalies are classified and located, and the anomaly detection data is output; S53, calculating the warning level based on the abnormality detection data, the state assessment data and the pre-stored warning threshold parameters to obtain warning level data; Obtain the current system context, generate decision suggestions based on the warning level data and the current system context, and obtain decision generation data; Based on anomaly detection data, status assessment data, pre-stored warning threshold parameters, warning level data and decision generation data, a system status report is generated, and ultimately the output includes status report data for operation and maintenance personnel, warning information data for the warning system and decision recommendation data for decision support.

7. A three-dimensional reconstruction system for power transmission lines based on a monocular channel monitoring camera, characterized in that: include: at least one processor; as well as, a memory communicatively connected to at least one of the processors; wherein, The memory stores instructions that can be executed by the processor, and the instructions are used to be executed by the processor to implement the three-dimensional reconstruction method of the transmission line based on the monocular channel monitoring camera as described in any one of claims 1 to 6.

Images