Rainy day substation intelligent patrol water drop influence monitoring and early warning method

Abstract

The application provides a rain day transformer substation intelligent patrol water drop influence monitoring and early warning method, and relates to the field of transformer substation intelligent operation and maintenance and state monitoring. The method provided by the application comprises the following steps: acquiring a double optical path video stream, constructing a cross-spectrum neural radiation field network, taking joint high-dimensional features as input, and outputting volume density and radiation color vectors; calculating equivalent water film thickness from the two, generating an equivalent water film thickness distribution field; mapping the equivalent water film thickness distribution field to a three-dimensional model, subdividing the grid according to the modulus of the thickness gradient, and solving the electric field to obtain the electric field distortion tensor and calibrate the high discharge risk critical region; generating a Lagrangian particle set, solving the dynamic equation through smooth particle hydrodynamics, simulating water film evolution to obtain the conductive water bridge through probability; inputting the distribution field, distortion tensor and through probability into a multi-modal impedance reconstructor to generate an electrochemical impedance spectrum to obtain a dynamic insulation margin decay rate; determining whether the condition triggers an early warning. The application significantly improves the timeliness and accuracy of rain day external insulation fault early warning.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance and condition monitoring of substations, and in particular to a method for monitoring and early warning of the impact of water droplets during intelligent inspections of substations in rainy weather. Background Technology

[0002] Against the backdrop of the rapid development of ultra-high voltage and smart grids, the safe operation of substation external insulation equipment in rainy weather is crucial. However, existing intelligent inspection and monitoring technologies have significant limitations and cannot meet the needs of high-precision early warning. First, mainstream non-contact visual monitoring methods have a bottleneck in sensing the surface. Current substation intelligent inspection systems mainly rely on visible light cameras combined with deep learning algorithms, which can only capture the macroscopic water droplet morphology and cannot detect the implicit wetting mechanisms such as the evolution of microscopic water films and ion migration. Especially in the early stages of drizzle, high humidity, or the dissolution of dirt, no obvious visible water flow has yet formed on the surface of the insulator, but the ion migration in the microscopic water film has already caused a sharp decline in insulation performance. The visual system is completely ineffective against this kind of implicit wetting, resulting in a serious delay in early warning. Anomalies are often only detected when the risk of flashover is extremely high.

[0003] Secondly, traditional contact-based electrical measurement methods face the dual challenges of narrow frequency domain coverage and weak anti-interference capabilities. Existing online monitoring devices mostly employ single-frequency or narrowband AC signal injection methods to measure leakage current or local impedance. Their operating frequencies are typically fixed at 50Hz or its harmonics, failing to cover the low-frequency band characterizing ion diffusion and the high-frequency band reflecting interface polarization, making it difficult to comprehensively assess insulation status. More seriously, in heavy rain environments, randomly falling water droplets, flowing water lines, and strong electromagnetic noise severely distort single-frequency measurement signals. Existing algorithms lack effective time-dependent capture and multi-modal feature fusion mechanisms, making it difficult to extract the true dielectric response characteristics from the strong noise background. This results in drastic fluctuations in monitoring data, a high false alarm rate, and an inability to provide stable and reliable quantitative data under complex and variable weather conditions. Therefore, it is urgent to develop a solution to these problems. Summary of the Invention

[0004] The purpose of this invention is to provide a method for intelligent inspection and monitoring of water droplet impact in substations during rainy weather, thereby improving the existing technical problems described in the background section.

[0005] The present invention provides a method for intelligent inspection and monitoring of water droplet impact in rainy weather substations, which adopts the following technical solution:

[0006] A method for intelligent inspection and monitoring of water droplet impact in rainy weather substations, specifically including:

[0007] A dual-path video stream, including a visible light video stream and an infrared video stream, is acquired using a substation camera during a rainy day. A cross-spectral neural radiation field network is constructed based on the dual-path video stream, and the joint high-dimensional features, formed by splicing the attribute information of the sampling points, are used as input to output the volume density and radiation color vector. The equivalent water film thickness along the observation ray is calculated based on the volume density and the radiation color vector to generate an equivalent water film thickness distribution field.

[0008] The equivalent water film thickness distribution field is mapped onto a three-dimensional model of the monitoring area. The grid is subdivided according to the thickness gradient modulus and the electric field is solved. The electric field distortion tensor is calculated, and the high discharge risk critical region is calibrated based on the eigenvalues ​​of the electric field distortion tensor.

[0009] A collection of Lagrange particles is generated within the water film region of the equivalent water film thickness distribution field. The dynamic equations of the coupled electric field force are solved by smooth particle hydrodynamics to simulate the evolution of the water film and calculate the probability of the conductive water bridge being connected.

[0010] The equivalent water film thickness distribution field, the electric field distortion tensor, and the conductive water bridge connection probability are input into the multi-mode impedance reconstructor to generate an electrochemical impedance spectrum. The dynamic insulation margin attenuation rate is calculated based on the electrochemical impedance spectrum. When the change in the dynamic insulation margin attenuation rate meets the preset conditions or the high discharge risk critical region exceeds the threshold, an early warning is triggered.

[0011] This invention provides a method for intelligent inspection and monitoring of water droplet impact in rainy substations. It constructs a cross-spectral neural radiation field network using visible and infrared dual-path video streams to reconstruct the equivalent water film thickness distribution field; solves the electric field distortion tensor based on the water film thickness gradient through adaptive subdivision of the grid; simulates the water film evolution and calculates the probability of conductive water bridge connection by coupling electric field forces with smooth particle hydrodynamics; and inputs multi-physics field characteristics into a multi-modal impedance reconstructor to calculate the dynamic insulation margin attenuation rate. This method enables monitoring and risk warning of the external insulation water film status of substations during rainy weather.

[0012] Optionally, when constructing a cross-spectral neural radiation field network based on the dual-optical-path video stream, and using the joint high-dimensional features formed by splicing sampling point attribute information as input, the output volume density and radiation color vector include:

[0013] The sampling point attribute information includes the spatial coordinates of the sampling point, the observation direction, and the infrared thermal radiation characteristics.

[0014] The spatial coordinates are mapped into spatial feature vectors through high-frequency sinusoidal position encoding, the observation direction is mapped into directional feature vectors through spherical harmonic function encoding, and the infrared thermal radiation features are mapped into thermal radiation feature vectors through a fully connected layer.

[0015] The spatial feature vector, the directional feature vector, and the thermal radiation feature vector are concatenated along the feature dimension to form the joint high-dimensional feature.

[0016] The joint high-dimensional features are input into the density head branch shallow neural network of the cross-spectral neural radiation field network. Only the spatial feature vectors in the joint high-dimensional features are received and compressed by two layers of linear transformation and high-dimensional features to output the volume density.

[0017] The joint high-dimensional features are input into the color head branch deep neural network of the cross-spectral neural radiation field network. The joint high-dimensional features are received and simulated through 7 layers of nonlinear transformation to output a radiation color vector.

[0018] Optionally, when calculating the equivalent water film thickness along the observed ray based on the volume density and the radiation color vector, and generating the equivalent water film thickness distribution field, the following steps are included:

[0019] Non-uniform sampling is performed on the observation ray that passes through the target pixel from the camera optical center to obtain the volume density of each sampling point and the infrared radiance in the radiative color vector.

[0020] The equivalent water film thickness is obtained by weighted integral of the volume density and the infrared radiance along the observation ray using the cross-spectral physical constraint density integral equation.

[0021] Based on the numerical value of the equivalent water film thickness, the image region is divided into different topological categories, forming an equivalent water film thickness distribution field that includes thickness values ​​and category labels.

[0022] Optionally, when mapping the equivalent water film thickness distribution field to a three-dimensional model of the monitoring area, subdividing the mesh according to the thickness gradient modulus, and solving for the electric field, the calculation of the electric field distortion tensor includes:

[0023] The equivalent water film thickness distribution field is assigned to the initial coarse mesh nodes of the three-dimensional model using nearest neighbor interpolation;

[0024] Calculate the thickness gradient magnitude of the water film thickness field in each grid cell, construct an error indicator based on the thickness gradient magnitude, and define a target grid side length function to drive adaptive grid subdivision.

[0025] On the subdivided grid, the dielectric constant tensor is calculated based on the real-time monitored ambient temperature and rainwater conductivity, and the dielectric constant tensor is assigned to the grid of the water film coverage area.

[0026] A nonlinear Poisson equation is constructed and the electric field distribution is solved by Newton-Raphson iteration. Based on the electric field distribution, the electric field distortion tensor is calculated for each point in the water film computational domain based on the dielectric constant tensor.

[0027] Optionally, when calibrating the high discharge risk critical region based on the eigenvalues ​​of the electric field distortion tensor, the following steps are included:

[0028] Calculate the maximum eigenvalue of the electric field distortion tensor; obtain the amplitude of the water film electric field at each point on the subdivided grid;

[0029] The region that simultaneously satisfies the condition that the amplitude of the water film electric field is greater than the preset air breakdown threshold and the maximum eigenvalue is greater than the preset distortion threshold is designated as a high discharge risk critical region.

[0030] Optionally, when generating a set of Lagrange particles within the water film region of the equivalent water film thickness distribution field, and solving the dynamic equations of the coupled electric field force through smooth particle hydrodynamics to simulate the water film evolution and calculate the probability of conductive water bridge connection, the process includes:

[0031] Within the water film region of the equivalent water film thickness distribution field, an initial smooth particle set is generated using an adaptive mesh mapping strategy, and the discretization mapping from image pixel data to Lagrange particles is completed through a smooth kernel function.

[0032] The particle acceleration is obtained by solving the discrete Navier-Stokes equations, and the particle velocity and position are updated by using a predictive-corrected symplectic integrator combined with dynamic smooth length adjustment.

[0033] The distribution changes of particles on the outer insulation surface at future moments are statistically analyzed to generate a wetting front velocity field. The system detects whether a connected path with a particle density continuously exceeding a preset threshold is formed between the high-voltage end and the grounding end. If such a path exists, the probability of the conductive water bridge being connected is calculated.

[0034] Optionally, the multimodal impedance reconstructor is obtained by training a neural network model based on the Transformer architecture.

[0035] Optionally, the equivalent water film thickness distribution field, the electric field distortion tensor, and the probability of conductive water bridge connection are input into a multimode impedance reconstructor to generate an electrochemical impedance spectrum. When calculating the dynamic insulation margin attenuation rate based on the electrochemical impedance spectrum, the following steps are included:

[0036] An initial high-dimensional feature vector is constructed, and principal component analysis is used to reduce the dimensionality of the initial high-dimensional feature vector in the computational domain of the three-dimensional model to obtain a low-dimensional multimodal feature vector.

[0037] The low-dimensional multimodal eigenvectors are input into the multimodal impedance reconstructor to obtain the electrochemical impedance spectrum;

[0038] The key resistance parameters were extracted by fitting the electrochemical impedance spectroscopy to the Landers equivalent circuit, and the dynamic insulation margin attenuation rate was calculated by Kalman filtering and differentiation.

[0039] Optionally, when the change in the dynamic insulation margin attenuation rate meets a preset condition or the high discharge risk critical region exceeds a threshold, the warning is triggered, including:

[0040] The slope of the dynamic insulation margin attenuation rate curve is monitored, and when the slope exceeds a preset abrupt change threshold, it is determined that the preset condition is met.

[0041] Monitor the area of ​​the high discharge risk critical region, or calculate the arc induction index based on the amplitude of the water film electric field and the maximum eigenvalue of the electric field distortion tensor within the high discharge risk critical region; when the area exceeds a preset area threshold or the maximum arc induction index exceeds a preset index threshold, it is determined to exceed the threshold.

[0042] An early warning signal is generated when any of the judgment conditions are met.

[0043] Optionally, it also includes:

[0044] The dynamic insulation margin attenuation rate, along with weather forecast data and real-time power grid load flow status, are input into a cleaning decision-making agent trained by reinforcement learning to generate a three-dimensional risk heat map and adaptive active intervention commands. Based on the adaptive active intervention commands, the cleaning equipment or dehumidification device is controlled to perform protective actions.

[0045] The training and decision-making process of the cleaning decision agent includes: using a generative adversarial network to generate a virtual extreme weather evolution scenario based on a historical extreme weather disaster database; driving the cleaning decision agent to perform Monte Carlo tree search in the virtual extreme weather evolution scenario; rehearsing and evaluating the effects of different prevention and control schemes; selecting the optimal prevention and control scheme; and generating the adaptive active intervention instruction based on the optimal prevention and control scheme.

[0046] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0047] This invention can penetrate rain to perceive the evolution mechanism of microscopic water films on the surface of external insulation, achieving precise quantification of dielectric response across the entire frequency band, from millihertz ion diffusion to kilohertz interface polarization. It effectively overcomes random raindrop morphology changes and environmental noise interference, significantly improving the robustness and accuracy of conductivity risk assessment of polluted water films under complex rain conditions. It shifts the early warning of insulation failure from macroscopic post-event observation to the early microscopic evolution stage, fundamentally solving the problems of delayed early warning and high false alarm rate in existing technologies, and providing a reliable basis for flashover prevention decisions in substations during rainy weather. Attached Figure Description

[0048] Figure 1 The flowchart illustrates a method for monitoring and early warning of water droplet impact during intelligent inspection of a substation in rainy weather, provided as an embodiment of the present invention. Detailed Implementation

[0049] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Unless otherwise defined, the technical or scientific terms used herein should have the ordinary meaning understood by those skilled in the art. The terms "comprising" and similar expressions used herein mean that the element or object preceding the word covers the element or object listed after the word and its equivalents, but does not exclude other elements or objects.

[0050] See Figure 1 This invention provides a method for intelligent inspection and monitoring of water droplet impact in rainy weather substations, comprising the following steps:

[0051] S1. Acquire dual-path video streams in rainy weather scenarios using substation cameras. The dual-path video streams include visible light video streams and infrared video streams. Construct a cross-spectral neural radiation field network based on the dual-path video streams, and use the joint high-dimensional features spliced ​​from the attribute information of sampling points as input to output volume density and radiation color vectors. Calculate the equivalent water film thickness along the observation ray based on the volume density and radiation color vectors to generate the equivalent water film thickness distribution field.

[0052] S2. Map the equivalent water film thickness distribution field to the three-dimensional model of the monitoring area, subdivide the grid according to the thickness gradient modulus and solve the electric field, calculate the electric field distortion tensor, and calibrate the high discharge risk critical area based on the eigenvalue of the electric field distortion tensor.

[0053] S3. Generate a set of Lagrange particles in the water film region of the equivalent water film thickness distribution field, solve the dynamic equation of the coupled electric field force through smooth particle hydrodynamics, simulate the evolution of the water film and calculate the probability of the conductive water bridge to be connected.

[0054] S4. Input the equivalent water film thickness distribution field, electric field distortion tensor, and conductive water bridge connection probability into the multi-mode impedance reconstructor to generate an electrochemical impedance spectrum. Calculate the dynamic insulation margin attenuation rate based on the electrochemical impedance spectrum. Trigger an early warning when the change in the dynamic insulation margin attenuation rate meets preset conditions or when the high discharge risk critical region exceeds a threshold.

[0055] In some embodiments, in step S1, a dual-optical-path video stream is acquired using a substation camera in a rainy scene. The dual-optical-path video stream includes a visible light video stream and an infrared video stream. When constructing a cross-spectral neural radiation field network based on the dual-optical-path video stream, and using the joint high-dimensional features formed by splicing sampling point attribute information as input, the output volume density and radiation color vector specifically include the following processes:

[0056] First, visible light and infrared video streams are simultaneously acquired using a dual-spectrum pan-tilt camera deployed near key equipment in the substation. This camera is mounted on a dedicated pole or equipment frame beam near the key equipment to obtain a stable and unobstructed viewing angle. The lens is directly facing the flashover-prone area of ​​the key equipment, at a distance of 3-8 meters. The key equipment being monitored includes four core external insulation components: transformer bushings, various insulator strings (suspension, post, tension), switchgear bushings, and surge arrester jackets. The visible light and infrared dual-path elements of the substation monitoring camera are simultaneously triggered to acquire frame-aligned video streams at a fixed frame rate. Based on a pre-calibrated dual-path intrinsic and extrinsic parameter matrix, geometric distortion correction is performed on each frame of the infrared image, and a depth mapping method is used to reproject the infrared image onto the visible light image coordinate system, ensuring strict spatial registration of the dual-path data.

[0057] Key equipment regions are identified in visible light images using a semantic segmentation network, generating a binarized mask. Each valid pixel within the binarized mask is traversed, and perspective projection effects are eliminated using the camera's intrinsic parameters. The pixel coordinates within the mask are then back-projected into normalized planar coordinates. Combined with extrinsic rotation and translation matrices, an observation ray originating from the camera's optical center and passing through the corresponding pixel's planar coordinates is constructed. The ray equation is:

[0058] ,

[0059] in, Let be the position function of the ray, representing the distance on the ray from the optical center. The coordinates of the three-dimensional point in space ; , , These are the specific position coordinates of the point on the three-dimensional horizontal horizontal axis, horizontal vertical axis, and vertical depth axis, respectively. The three-dimensional position coordinates of the camera's optical center in the world coordinate system are determined by the translation vector in the extrinsic parameters; The ray direction vector represents the observation direction from the camera's optical center to this pixel. The starting depth for sampling represents the nearest distance threshold for starting a valid data search along the ray direction, used to skip air regions without targets in front of the lens to improve computational efficiency. The sampling termination depth represents the farthest distance threshold along the ray direction where data searching stops. It is used to limit the maximum expected depth of the target device and exclude interference from infinitely distant background.

[0060] exist A non-uniform adoption strategy is adopted within the range: with a larger step size. Quickly locate the approximate position of the equipment's external insulation surface, near the interface where a water film or contaminant layer is expected to exist, using small step sizes. Sampling was performed to focus on details of the water film interface.

[0061] Secondly, for each sampling point, its three-dimensional coordinates are reprojected onto the visible light and infrared image planes using the ray casting principle. The corresponding RGB color values ​​and original infrared grayscale values ​​are then obtained through bilinear interpolation. Based on the radiometric calibration parameters of the infrared camera, the infrared grayscale values ​​are converted into relative thermal radiation intensity.

[0062] The three-dimensional spatial coordinates of each sampling point are mapped into an anti-aliasing spatial feature vector using high-frequency sinusoidal position encoding, and the ray direction vector is encoded into a directional feature vector using a spherical harmonic function. Simultaneously, the relative thermal radiation intensity is mapped into a thermal radiation feature vector through a small fully connected layer. The output dimension of this fully connected layer is a configurable hyperparameter, and its value (e.g., 8, 16, 32) depends on the feature fusion mechanism's requirement for expressing thermophysical information. These three types of feature vectors are then concatenated and fused along their respective feature dimensions to form a joint high-dimensional feature for the sampling point.

[0063] Finally, high-dimensional features are input into a cross-spectral neural radiation field network, which consists of a multilayer perceptron sharing low-level feature extraction, with its deep branches solving two key physical fields in parallel:

[0064] First, the density head branch is a shallow neural network that only receives the spatial feature vector from the joint high-dimensional features. After two layers of linear transformation and high-dimensional feature compression, it outputs the volume density, which is used to accurately characterize the probability of visible light scattering and absorption by the water medium at the sampling point.

[0065] Second, the color head branch is a deep neural network that receives complete joint high-dimensional features and simulates complex physical optical processes through seven layers of nonlinear transformation. This branch uses spatial and directional feature vectors to learn Fresnel reflection and specular highlighting effects in the visible light band, i.e., the change in the intensity of water surface reflection at different angles; simultaneously, it uses thermal radiation and directional feature vectors to learn the emissivity anisotropy in the infrared band, i.e., the minute changes in the thermal radiation intensity of the water film itself at different observation angles. The network outputs a radiative color vector, where the first three channels correspond to the three primary color components of the visible light band, and the fourth channel corresponds to the radiance value of the infrared band.

[0066] During the network training and optimization phase, in addition to the conventional photometric reprojection loss, a water film continuity regularization term is introduced into the loss function. This regularization term forces a constraint on the spatial rate of change of the density field by calculating the second-order difference of the density gradient between adjacent sampling points along the ray. It utilizes the natural smoothness prior caused by the surface tension of the liquid to effectively suppress density artifacts caused by noise or sparse viewpoints, ensuring that the reconstructed water film interface is topologically continuous and conforms to hydrodynamic properties.

[0067] In some embodiments, when calculating the equivalent water film thickness along the observed ray based on the volume density and the radiation color vector in step S1, and generating the equivalent water film thickness distribution field, the specific process includes the following:

[0068] First, for each observation ray passing through the target pixel from the camera's optical center, within the effective depth range Non-uniform sampling is performed to obtain a series of sampling points distributed along the ray. For each sampling point, the volume density value output by the cross-spectral neural radiation field network is queried, and the infrared radiance value of that point is extracted from the fourth channel of the radiation color vector.

[0069] Secondly, based on the volume density and radiation color vector output by the cross-spectral neural radiation field network, the equivalent water film thickness and its topological classification label corresponding to the pixel are calculated by integrating along the observation ray using the cross-spectral physical constraint density integral formula. This leverages the complementary thermal characteristics of liquid evaporation and cooling to address the detection blind zone of transparent water films under visible light, achieving accurate reconstruction of the three-dimensional morphology of water films on equipment surfaces under complex meteorological conditions. The formulas used are as follows:

[0070] ,

[0071] in, Pixel coordinates of planar images captured and registered by the camera The equivalent water film thickness at the location is expressed in micrometers. These are the horizontal and vertical pixel coordinates, respectively. ρ is the physical density constant of water; The dynamic confidence weights are determined based on the signal-to-noise ratio of the visible light image in the current rainy weather environment. When visible light images are severely affected by rain and have a low signal-to-noise ratio, Automatic reduction makes subsequent calculations more reliant on infrared thermal gradient information; when the visible light image is clear and the signal-to-noise ratio is high, The automatic scaling makes the calculation results primarily dependent on visible light geometric density information. for Sampling points within the range The volume density at that location; for Sampling points within the range Temperature gradient in the infrared band at that location; Sampling points Radiance value; for Reference value for the maximum temperature gradient modulus within the range; The cumulative transmittance term simulates light rays from position... Arrival Location The probability that it was not previously absorbed or scattered by the preceding medium, among which These are dummy variables on the integration path, representing intermediate points on the path; The cross-spectral weighted density represents the equivalent water film medium opacity that integrates optical and thermal characteristics; For the ray integral path and Between positions The cross-spectral weighted density at the location.

[0072] Finally, based on the calculated equivalent water film thickness, the water film regions in the two-dimensional planar image are classified: regions with a thickness less than ten micrometers and uniform distribution are labeled as monolayer adsorption regions; regions with a thickness greater than twenty micrometers and discrete cluster distribution are labeled as multilayer stacked regions; and elongated, thick regions with obvious flow gradients and continuity are labeled as flow marks. The final output is an equivalent water film thickness distribution field containing continuous thickness values ​​and discrete topological labels, providing accurate initial state data for subsequent electric field distortion assessment and water film evolution prediction.

[0073] In some embodiments, when mapping the equivalent water film thickness distribution field to a three-dimensional model of the monitoring area in step S2, subdividing the mesh according to the thickness gradient modulus and solving for the electric field, calculating the electric field distortion tensor, and calibrating the high discharge risk critical region based on the eigenvalues ​​of the electric field distortion tensor, the specific process includes the following:

[0074] First, oblique photogrammetry was used to scan the key monitoring areas of the substation to acquire point cloud data, and a high-fidelity 3D model containing real geometric defects was reconstructed. A nearest-neighbor interpolation algorithm was employed to discretize the continuous equivalent water film thickness distribution field data and assign it to each node of the initial coarse grid of the 3D model, thereby establishing a mapping relationship between the equivalent water film thickness distribution field and the surface of the 3D model. For dry areas in the water film thickness distribution field that are not covered by a water film, the water film thickness was forcibly set to zero to ensure a clear boundary for the continuity of the water film.

[0075] Based on the initial mesh, the thickness gradient magnitude at the center and nodes of each element is accurately calculated using the superconvergence recovery technique. Subsequently, an error indicator was constructed as the sole driving force for the mesh density distribution:

[0076] ,

[0077] in, For the first The comprehensive estimation error of the water film thickness field within each grid cell; the larger the value, the more the grid in that area needs to be refined. These are the gradient and curvature weight coefficients, respectively, and their sum is one. For the first Water film thickness within each unit The gradient magnitude; Indicates the first Water film thickness field within each unit The local curvature.

[0078] Based on an error indicator, a target mesh edge length function is defined to control the size and shape of the mesh, causing the mesh edge length to automatically decrease as the gradient increases. In high-gradient regions such as the water film tip and water bridge neck, the mesh is forcibly refined to the millimeter level, while maintaining a sparser mesh in regions of uniform thickness. This allows for the resolution of elongated watermark edges using a minimum number of elements. The formula used is as follows:

[0079] ,

[0080] in, In spatial location The characteristic size of the mesh cells to be generated; This represents the coarsest grid size when the water film thickness changes gradually. This represents the finest mesh size when the water film thickness changes drastically. The gradient threshold is used as a reference to determine the critical point at which the mesh begins to be significantly refined. As a transition index, it controls the grid size from Transition to The rate; min() is the truncation operator, ensuring that the calculated target side length will never exceed the preset maximum value. .

[0081] Mesh reconstruction is performed using a frontier advance method combined with the Delaunay kernel splitting algorithm: traversing the current mesh of the 3D model and marking all... Cells exceeding the error threshold are considered cells to be refined, especially at water bridge connections and the tip of the water film. A recursive binary search is performed on these cells, and the newly generated node coordinates strictly follow the target mesh edge length function. Constraints were applied, and the Laplace smoothing algorithm was used to adjust the new node positions to ensure that the element quality met the stability requirements of the finite element solution, namely, a minimum interior angle greater than 15 degrees and a maximum aspect ratio less than 10. The original equivalent water film thickness distribution data was mapped onto the newly generated adaptive mesh. Typical high field strength regions were selected to compare the rate of change of electric field strength before and after densification. If the relative error was less than 1%, the mesh density was determined to have reached millimeter-level accuracy; otherwise, the mesh side length function was returned to further reduce the density. And then re-divided.

[0082] Secondly, the ambient temperature is monitored in real time by high-precision digital temperature sensors deployed in key monitoring areas and dedicated rainwater collection and conductivity measurement devices. and rainwater conductivity The complex permittivity tensor of rainwater is calculated using an empirical formula:

[0083] ,

[0084] in, The real part is taken as the static dielectric constant of pure water; the imaginary part... middle The imaginary unit indicates that the imaginary part has a 90-degree phase difference with the real electric field, representing the energy loss portion; It is the power frequency angular frequency; is the vacuum permittivity.

[0085] The above dielectric constant tensor is used to assign values ​​to the mesh of the three-dimensional model: for air meshes, the dielectric constant tensor is taken as the vacuum dielectric constant. For the mesh of the external insulation surface of the equipment, the dielectric constant tensor matrix is:

[0086] ,

[0087] in, Location of the outer insulation surface The dielectric constant tensor matrix at that location; is the relative permittivity of the external insulating surface material along the fiber direction, which is determined by the fiber material itself, has a small value, and takes a range of [3,4]. The relative permittivity of the material perpendicular to the fiber direction is determined by the matrix resin and has a relatively large value, ranging from [5, 6.5]. It is the identity matrix; A column vector describing the local fiber orientation; This is the transpose of the column vector, i.e., the row vector. The dielectric constant tensor of the raindrop grid region is... .

[0088] Subsequently, in the finite element model solver, the mesh is reconstructed based on the 3D model, and the nonlinear Poisson equation is constructed:

[0089] ,

[0090] The solution domain of this equation is the three-dimensional model region; For divergence operators; The location within the 3D model area is The dielectric constant tensor; electric potential The rate of change of space, i.e., the potential gradient; Space charge density; electric potential and electric field These are two key physical parameters that determine the distribution of space charge density.

[0091] Initially set space charge density If the value is zero, a nonlinear coupled iterative loop is performed using the Newton-Raphson method. In each iteration, the solver adjusts the value based on the potential of the current grid node. and the local electric field derived from its gradient The space charge density of each grid point is dynamically updated by calling the preset bipolar drift-diffusion equation. This leads to the construction of the Jacobian matrix and residual vector of the nonlinear Poisson equation reflecting the latest charge distribution. By linearizing and solving the large sparse linear system of equations formed after discretization of the Poisson equation within the computational domain, the potential correction is obtained. The potential distribution is updated globally, and the residual norm is checked to see if it is below the set convergence tolerance. If convergence is not achieved, the above process is repeated until the potential field and space charge distribution reach a self-consistent equilibrium.

[0092] Using the converged high-precision potential, gradient calculations are performed on the unit nodes, and an interpolation algorithm is used to generate a global local electric field distribution map containing precise amplitude and three-dimensional direction vector, thus fully revealing the microscopic influence of space charge accumulation on electric field distortion.

[0093] Finally, based on the above global and local electric field distribution map and the equivalent water film thickness distribution field, the electric field distortion tensor is calculated:

[0094] ,

[0095] in, The electric field distortion tensor is a symmetric matrix used to quantitatively characterize the intensity and direction of local electric field distortion caused by the combined effect of non-uniform water film and fouling layer. This refers to the computational domain of the water film in the 3D model. It is a volume infinitesimal element; Let be the dielectric constant tensor of rainwater; Represents the gradient of dielectric constant; The dielectric constant gradient magnitude; To prevent dividing by zero and minimizing positive numbers. In the formula... The term approaches 0 inside a uniform region, indicating that the interior of the uniform medium does not contribute to the distortion; it approaches 1 when the dielectric constant changes drastically at the interface between rainwater and air, ensuring that only the interface participates in the distortion calculation. The attenuation coefficient is a value greater than 0. The larger the value, the faster the weight decays in regions far from the electrode; For the current point Euclidean distance to the surface of the high-voltage electrode; For position Water film thickness at the location; Dry location The operating electric field can be obtained from the local electric field distribution map of the entire region. For tensor product, It is a second-order tensor matrix that preserves the direction information of the electric field and the energy density distribution.

[0096] Solve for the electric field distortion tensor Characteristic equation Three eigenvalues ​​were obtained. , respectively representing points The distortion energy density along the three principal axes of the electric field distortion is taken as the largest eigenvalue. .

[0097] Combined with the amplitude of the water film electric field Regions that simultaneously meet the following two conditions are selected as high-discharge-risk critical regions. :

[0098] and ,

[0099] in, For position The amplitude of the electric field of the water film at that location; The air breakdown threshold; For position The largest eigenvalue at; The distortion significance threshold is set based on experience. The calibration results of this critical region provide a precise spatial positioning basis for subsequent early warning trigger judgment.

[0100] In some embodiments, when generating a set of Lagrange particles in the water film region of the equivalent water film thickness distribution field in step S3, and solving the dynamic equation of the coupled electric field force through smooth particle hydrodynamics to simulate the evolution of the water film and calculate the probability of the conductive water bridge being connected, the specific process includes the following:

[0101] First, based on the water film region in the equivalent water film thickness distribution field identified in step S1, an initial smooth particle set is generated within the water film domain using an adaptive mesh mapping strategy. This ensures that the particle number density is proportional to the local water film thickness, thus maintaining mass conservation. The smooth particles discretize the continuous water film fluid into a set of Lagrangian particles carrying physical properties, naturally adapting to the complex topological changes that occur in water droplets under electric field driving, such as large deformations, free surface evolution, breakup, and merging.

[0102] Each particle is assigned physical properties, including rest mass, initial velocity, initial pressure, and surface tension coefficient of water, determined by the particle spacing and water density. The initial velocity is directly mapped from the pixel-level displacement vector field calculated by the optical flow method of the current image frame, extending the two-dimensional optical flow into a three-dimensional velocity and adding small random perturbations to simulate thermal fluctuations. The initial pressure is initialized based on the hydrostatic pressure distribution.

[0103] The local summation density of the initial particles is calculated using the smooth kernel function in the smooth particle hydrodynamics algorithm and compared with the theoretical water density. If the deviation exceeds a preset threshold, the initial particle distribution is corrected by removing overlapping particles or inserting new particles in low-density regions, thus completing the discretization mapping from image pixel data to a Lagrange particle system with complete dynamic properties.

[0104] Secondly, within each time step, based on the current particle distribution configuration, the local density of a spherical neighborhood centered on the target particle with a radius equal to the smooth length is estimated using a smooth kernel function, and the pressure is calculated by solving the weakly compressible state equation. Then, the pressure gradient is obtained by using the particle approximation method.

[0105] A discretized Navier-Stokes equation is constructed, where the left-hand side represents the particle acceleration to be solved, and the right-hand side consists of pressure gradient, viscosity, and external force terms. The viscosity term quantifies the shear stress diffusion effect caused by the velocity gradient within the fluid by performing a second spatial differentiation of the particle velocity field. The external force terms consist of particle gravity, dielectrophoresis, aerodynamic drag, and surface tension. The particle gravity term is obtained by multiplying the particle's mass by a constant gravitational acceleration; the dielectrophoresis term is derived by calculating the square of the local electric field gradient in a spherical shape, combined with the difference in dielectric constant between water and air and the particle volume; the aerodynamic drag term is calculated based on the relative difference between the particle velocity and the background air velocity, substituted into Stokes' law; and the surface tension term is obtained by converting the curvature and tension coefficient at the interface into equivalent volume forces using a continuous surface force model, coupled with a dynamic contact angle model for correction. Solving the discrete Navier-Stokes equation yields the instantaneous acceleration of each particle.

[0106] Based on the particle acceleration obtained from the solution, a predictive-correction symplectic integrator is used to advance the particle's kinematic state: in the prediction step, the intermediate velocity and intermediate position of the particle are calculated, and in the correction step, the resultant force is introduced to update the final velocity and position.

[0107] Simultaneously, the inter-particle interaction logic is executed: the particle pairs are traversed and the detection distance is checked. If the distance is less than the preset merging threshold, the merging algorithm is triggered to merge the mass, momentum, and charge of the two particles into a single particle according to the conservation principle, in order to simulate the water droplet fusion process. For the solid wall boundary of the insulating jacket, the collision mode of the particles is determined according to the preset wettability parameters: when the adhesion force is less than the inertial force, the particles undergo elastic or inelastic collision and rebound; when the adhesion force is greater than the inertial force, the particles switch to the adhesion and sliding mode and slowly migrate along the wall surface.

[0108] Furthermore, the smoothing length hh is dynamically adjusted based on the real-time changes in particle number density to maintain computational accuracy and numerical stability. Through the above process, the dynamics of raindrops within this time step are fully realized, and the updated velocity and latest position of each particle are obtained.

[0109] Finally, based on the updated particle position and velocity, the distribution changes of particles on the outer insulating surface at future time points are statistically analyzed. A formula for calculating the wetting front flux is constructed:

[0110] ,

[0111] in, Position in the 3D model At the current moment The wetting front flux vector represents the net mobility of the water film mass on the outer insulation surface in the spatial grid in the three-dimensional model. For inertial convection; For position At the current moment The water flow velocity is the instantaneous velocity at the current moment, which is obtained by integrating the calculated particle acceleration over time. The macroscopic flow velocity at the center of the grid is obtained by weighted averaging of the velocities of all particles within the 3D model grid. For tension diffusion term, For divergence operators; The dynamic surface tension coefficient is determined by the dynamic contact angle. The result is obtained by substituting the modified Young's equation into the function mapping calculation, which reflects the change of interfacial energy with the wetting state. It describes the dynamic contact angle at the contact line between the water, solid, and gas phases; The interface curvature drives the surface tension flow, which is calculated using differential geometric formulas based on the water film thickness distribution data and its spatial gradient. Unit tensor; The interface normal vector is obtained by constructing a vector perpendicular to the tangent plane by calculating the spatial gradient of the height of the water-air interface, and then normalizing its length to a unit length. For the normal projection tensor; This is the transpose of the interface normal vector. For dielectric driving term, Dielectric mobility is calculated by dividing the experimentally measured conductivity of rainwater by the product of the carrier concentration and the elementary charge. Let be the local electric field intensity, which is the vector value at position r in the global local electric field distribution map; The relative permittivity gradient describes the spatial rate of change of dielectric properties at the water-air interface. This is the term for thin film friction damping. This is a correction factor, with a value greater than 1, used to correct the external insulation surface resistance calculated by the basic Heron-Shore flow model, making it closer to the actual resistance; The aerodynamic viscosity is obtained by consulting the standard physical property table to find the absolute viscosity value of air at the corresponding temperature. The thickness of the water film.

[0112] Divide the flux vector of the wetting front by the surface density of the local water film to calculate the macroscopic velocity field. Extract the component of this velocity field in the normal direction of the water film boundary as the propulsion rate of the wetting front, and use this rate to drive the position update of particles, thus presenting the dynamic evolution of the water film boundary within the future time step.

[0113] Within a preset future time window, the system continuously monitors whether a connected path with a particle density continuously exceeding a preset threshold is formed between the high-voltage end and the grounding end. If such a connected path exists, the probability of a conductive water bridge is calculated based on the continuous length and density integral of the path to quantify the flashover risk. The output of this probability value provides crucial dynamic prediction basis for subsequent early warning triggering judgments and proactive intervention decisions.

[0114] In some embodiments, in step S4, the equivalent water film thickness distribution field, electric field distortion tensor, and the probability of conductive water bridge connection are input into the multi-mode impedance reconstructor to generate an electrochemical impedance spectrum. The dynamic insulation margin attenuation rate is calculated based on the electrochemical impedance spectrum. When the change in the dynamic insulation margin attenuation rate meets a preset condition or the high discharge risk critical region exceeds a threshold, an early warning is triggered. Specifically, the following process is included:

[0115] First, three physical quantities are collected from the mesh nodes within the computational domain of the current 3D model: water film thickness, eigenvalues ​​of the electric field distortion tensor, and the probability of a conductive water bridge connecting. To avoid the dominant features being masked by dimensional differences, these three physical quantities are normalized separately. The processed data are then concatenated according to their spatial positions in the 3D model to construct an initial high-dimensional feature vector.

[0116] Based on the initial high-dimensional spatiotemporal eigenvectors of all meshes in the 3D model, the covariance matrix of these eigenvectors is calculated, and the principal component directions are obtained through eigenvalue decomposition. The principal component directions corresponding to the top g largest eigenvalues ​​are selected as the projection matrix based on the cumulative variance contribution rate threshold. The initial high-dimensional spatiotemporal eigenvectors are multiplied by this projection matrix to complete the dimensionality reduction mapping, outputting a low-dimensional multimodal eigenvector containing minimal redundant information and preserving the key driving factors of leakage current to the greatest extent possible.

[0117] Secondly, the low-dimensional multimodal feature vectors are serialized, and positional encoding is added before inputting them into a pre-trained multimodal impedance reconstructor. This reconstructor is a neural network model based on the Transformer architecture, which utilizes its multi-head self-attention mechanism to capture the complex nonlinear coupling and temporal dependencies between features. Internally, the model performs deep mapping inference through a feedforward neural network and a nonlinear activation function to directly regress the complex impedance spectrum in the output frequency domain.

[0118] This electrochemical impedance spectroscopy (EIS) covers a wide bandwidth from millihertz to kilohertz: low-frequency millihertz data characterize ion diffusion processes within thick water films, while high-frequency kilohertz data characterize rapid charge-discharge processes at the solid-liquid interface. The discrete frequency data output from the model are smoothed and interpolated, and their physical consistency is verified to reconstruct a complete and highly accurate EIS, reflecting the dielectric response characteristics of the external insulating surface under wet conditions in real time.

[0119] Then, the electrochemical impedance spectroscopy data were fitted with an equivalent circuit using the nonlinear least squares method, employing the Landels equivalent circuit model. This model includes solution resistance, charge transfer resistance, double-layer capacitance, and Warburg impedance elements. Through fitting, the solution resistance, characterizing the conductivity of the fouling liquid film, and the charge transfer resistance, characterizing the interfacial polarization resistance, were accurately extracted from the electrochemical impedance spectroscopy as key state parameters.

[0120] Subsequently, the extracted resistance parameter sequence is smoothed and differentiated using Kalman filtering based on a sliding time window, and the dynamic insulation margin decay rate is calculated to quantify the rate of insulation performance loss per unit time. Finally, by monitoring the abrupt change in the slope of the dynamic insulation margin decay rate curve, the inflection point of the sudden drop in resistance caused by rapid dissolution of the contaminant layer and a surge in ion concentration is automatically identified, thereby providing an early warning of the critical moment when the external insulation surface develops from localized wetting to overall flashover.

[0121] In addition, based on the dynamic insulation margin attenuation rate and the state of the high discharge risk critical region, the following early warning judgment logic is executed:

[0122] Monitor the change of the dynamic insulation margin attenuation rate curve over time. When the slope of the attenuation rate curve exceeds the preset abrupt change threshold, it indicates that the insulation resistance is decreasing rapidly, the contaminant layer is dissolving rapidly, or the ion concentration is surging, and it is determined that the preset condition is met.

[0123] Monitor the state of the high-discharge-risk critical region. Calculate the area of ​​the critical region, or based on the amplitude of the water film electric field at each point within the critical region. With the largest eigenvalue of the electric field distortion tensor The arc-induced potential index is calculated using the following formula:

[0124] ,

[0125] in, For position The amplitude of the electric field of the water film at that location; The sensitivity coefficient is set to 2 to amplify the effect of high field strength. For position The largest eigenvalue at that location. There is no risk of discharge. This indicates the possibility of a weak partial discharge. This indicates a very high probability of flashover or arc bridging.

[0126] When either of the above two judgment conditions is met, the system generates an early warning signal and triggers the substation operation and maintenance alarm system.

[0127] In some preferred embodiments, after an early warning is triggered, the system further executes an active intervention decision-making process. This process constructs a clean decision-making agent based on a deep reinforcement learning framework. By fusing multi-source heterogeneous data, designing a hierarchical weighted composite reward function, generating extreme weather scenario simulations, and performing Monte Carlo tree search, it ultimately generates adaptive active intervention instructions and constructs a three-dimensional risk heat map.

[0128] Specifically, the cleaning decision unit acquires three types of data in real time: the dynamic insulation margin attenuation rate of the outer insulation surface, high-resolution gridded weather forecast data, and real-time power grid load flow status. The high-resolution gridded weather forecast data is obtained by accessing a numerical weather prediction system and integrating measured data from local micro-meteorological monitoring stations, processed using bilinear interpolation and bias correction algorithms, resulting in a gridded dataset containing temperature, humidity, rainfall, and pollution deposition rate. The real-time power grid load flow status is calculated in real time by collecting high-frequency measurement data from synchronous phasor measurement units and smart meters through a wide-area measurement system, followed by cleaning and topology analysis using a state estimation algorithm. After denoising, missing value imputation, and Minkowski distance normalization, these three types of data are concatenated to construct a reinforcement learning state space.

[0129] Specifically, a hierarchical weighted composite reward function is designed, which consists of a flashover penalty term, an energy cost term, and an equipment lifespan depreciation term. The flashover penalty term applies an exponentially high negative reward when the predicted flashover probability exceeds a threshold, thus establishing an insurmountable safety red line. The energy cost term calculates a negative reward based on the cleaning equipment power and operation time linearly, to suppress over-cleaning. The equipment lifespan depreciation term quantifies equipment wear penalties based on an action frequency and mechanical stress model. By dynamically adjusting the weight coefficients of each component, the cleaning decision-making agent is guided to automatically balance the three objectives of sudden disaster prevention, operational economy, and equipment lifespan maintenance through massive trial and error during deep reinforcement learning training, ultimately converging to the optimal adaptive cleaning strategy that takes into account multi-dimensional constraints.

[0130] Specifically, to enhance the agent's ability to cope with extreme operating conditions, a high-fidelity virtual simulation environment is constructed using a generative adversarial network (GAN). The GAN uses a historical extreme weather disaster database as its training sample set, a power grid equipment physical characteristic map as its constraint, and a real-time power grid topology and load profile as its initial state to simulate and generate the evolution of extreme rainstorm weather scenarios. The historical extreme weather disaster database includes minute-level rainfall, visibility, relative humidity, wind speed and direction, temperature change curves, and pollution growth rates for similar extreme rainstorm weather events that occurred in the target area over the past ten to twenty years. The power grid equipment physical characteristic map includes flashover voltage characteristic curves of insulation surfaces under different levels of pollution and humidity, a power-temperature response model of a heating and dehumidifying device, and a time-dependent decay model after water-repellent agent spraying. The real-time power grid topology and load profile includes the current power grid wiring configuration, real-time load distribution at each node, and reserve capacity distribution.

[0131] Specifically, in the evolutionary scenario, a closed-loop feedback mechanism is established, in which the evolution of the meteorological microenvironment drives the wetting of the insulation surface, the degradation of insulation performance triggers the redistribution of power flow in the power grid, and control actions reversely correct the equipment state. The driving cleaning decision agent takes the current power grid state as the root node and performs a Monte Carlo tree search for actions such as heating and dehumidification, load transfer, spraying water-repellent agents, and their combinations. The Monte Carlo tree search unfolds through an iterative process of four steps: selection, expansion, simulation, and backtracking. In the selection phase, starting from the root node, nodes to be expanded are selected downwards along the current optimal path according to the tree strategy; in the expansion phase, one or more legal actions are added as new child nodes to the selected leaf nodes; in the simulation phase, starting from the newly expanded nodes, the simulation is quickly extrapolated to the preset future 30-minute time endpoint based on a random or heuristic strategy, and indicators such as flashover risk, energy consumption cost, and equipment wear under this path are evaluated; in the backtracking phase, the simulation results are propagated backwards along the search path to update the access count and value estimate of each node. By evaluating the performance of different strategy combinations in a virtual simulation over the next 30 minutes through large-scale sampling, the Pareto optimal prevention and control scheme that takes into account the triple objectives of safety, economy and lifespan is finally selected, realizing the pre-simulation and dynamic optimization of extreme weather response strategies.

[0132] Specifically, the optimal prevention and control scheme selected through Monte Carlo tree search iterations is deeply analyzed and transformed into specific control commands issued to the physical execution end to initiate proactive defense actions, such as activating heating and dehumidification devices, adjusting load distribution, or triggering automatic cleaning equipment. Simultaneously, the massive state trajectory data generated during the Monte Carlo tree search simulation is retrospectively utilized, combined with meteorological evolution scenarios and power grid equipment physical characteristic maps, to quantitatively predict the probability of flashover failures in each monitoring area within future time windows due to pollution accumulation, humidification processes, and electric field distortion. The calculated probability values ​​are mapped to the power grid geographic information coordinate system, and a three-dimensional risk heat map is constructed using spatiotemporal interpolation algorithms. This three-dimensional risk heat map can intuitively present the migration path and intensity change trend of high-risk areas and dynamically display the expected risk reduction effect after the proposed control measures are taken. By comparing the differences in risk heat maps before and after the implementation of control measures, maintenance personnel can clearly assess the effectiveness of the current strategy. This visualization mechanism upgrades traditional static threshold alarms to forward-looking decision support based on dynamic prediction, enabling precise and proactive command of substation pollution flashover prevention work under rainy weather conditions.

[0133] While embodiments of the present invention have been described in detail above, it will be apparent to those skilled in the art that various modifications and variations can be made to these embodiments. However, it should be understood that such modifications and variations fall within the scope and spirit of the invention as set forth in the claims. Furthermore, the invention described herein may have other embodiments and can be implemented or carried out in various ways.

Claims

1. A method for intelligent inspection and monitoring of water droplet impact in rainy weather substations, characterized in that, include: The video stream is captured by a substation camera in a rainy weather scenario using a dual-optical-path video stream, which includes a visible light video stream and an infrared video stream. A cross-spectral neural radiation field network is constructed based on the dual-optical-path video stream, and the joint high-dimensional features formed by splicing sampling point attribute information are used as input to output volume density and radiation color vectors. Based on the volume density and radiation color vectors, the equivalent water film thickness along the observation ray is calculated, generating an equivalent water film thickness distribution field. Specifically, when constructing the cross-spectral neural radiation field network based on the dual-optical-path video stream and using the joint high-dimensional features formed by splicing sampling point attribute information as input to output volume density and radiation color vectors, the following is included: the sampling point attribute information includes the spatial coordinates of the sampling point, the observation direction, and infrared thermal radiation characteristics; the spatial coordinates are mapped into a spatial feature vector through high-frequency sinusoidal position encoding, and the observation direction is mapped into a spatial feature vector. The infrared thermal radiation features are mapped to directional feature vectors via spherical harmonic function encoding. These features are then mapped to thermal radiation feature vectors via fully connected layers. The spatial feature vectors, directional feature vectors, and thermal radiation feature vectors are concatenated along their feature dimensions to form the joint high-dimensional feature. This joint high-dimensional feature is then input into the density head branch of the cross-spectral neural radiation field network's shallow neural network. This network receives only the spatial feature vector from the joint high-dimensional feature and performs two layers of linear transformation and high-dimensional feature compression to output volume density. Finally, the joint high-dimensional feature is input into the color head branch of the cross-spectral neural radiation field network's deep neural network. This network receives the joint high-dimensional feature and performs seven layers of nonlinear transformation to simulate physical optical processes, outputting a radiation color vector. The equivalent water film thickness distribution field is mapped onto a three-dimensional model of the monitoring area. The grid is subdivided according to the thickness gradient modulus and the electric field is solved. The electric field distortion tensor is calculated, and the high discharge risk critical region is calibrated based on the eigenvalues ​​of the electric field distortion tensor. A collection of Lagrange particles is generated within the water film region of the equivalent water film thickness distribution field. The dynamic equations of the coupled electric field force are solved by smooth particle hydrodynamics to simulate the evolution of the water film and calculate the probability of the conductive water bridge being connected. The equivalent water film thickness distribution field, the electric field distortion tensor, and the conductive water bridge connection probability are input into the multi-mode impedance reconstructor to generate an electrochemical impedance spectrum. The dynamic insulation margin attenuation rate is calculated based on the electrochemical impedance spectrum. When the change in the dynamic insulation margin attenuation rate meets the preset conditions or the high discharge risk critical region exceeds the threshold, an early warning is triggered.

2. The method as described in claim 1, characterized in that, Calculating the equivalent water film thickness along the observed ray based on the volume density and the radiation color vector, and generating the equivalent water film thickness distribution field, includes: Non-uniform sampling is performed on the observation ray that passes through the target pixel from the camera optical center to obtain the volume density of each sampling point and the infrared radiance in the radiative color vector. The equivalent water film thickness is obtained by weighted integral of the volume density and the infrared radiance along the observation ray using the cross-spectral physical constraint density integral equation. Based on the numerical value of the equivalent water film thickness, the image region is divided into different topological categories, forming an equivalent water film thickness distribution field that includes thickness values ​​and category labels.

3. The method as described in claim 1, characterized in that, Mapping the equivalent water film thickness distribution field onto a three-dimensional model of the monitoring area, subdividing the mesh based on the thickness gradient modulus and solving for the electric field, and calculating the electric field distortion tensor include: The equivalent water film thickness distribution field is assigned to the initial coarse mesh nodes of the three-dimensional model using nearest neighbor interpolation; Calculate the thickness gradient magnitude of the water film thickness field in each grid cell, construct an error indicator based on the thickness gradient magnitude, and define a target grid side length function to drive adaptive grid subdivision. On the subdivided grid, the dielectric constant tensor is calculated based on the real-time monitored ambient temperature and rainwater conductivity, and the dielectric constant tensor is assigned to the grid of the water film coverage area. A nonlinear Poisson equation is constructed and the electric field distribution is solved by Newton-Raphson iteration. Based on the electric field distribution, the electric field distortion tensor is calculated for each point in the water film computational domain based on the dielectric constant tensor.

4. The method as described in claim 1, characterized in that, When calibrating the high discharge risk critical region based on the eigenvalues ​​of the electric field distortion tensor, the following is included: Calculate the maximum eigenvalue of the electric field distortion tensor; obtain the amplitude of the water film electric field at each point on the subdivided grid; The region that simultaneously satisfies the condition that the amplitude of the water film electric field is greater than the preset air breakdown threshold and the maximum eigenvalue is greater than the preset distortion threshold is designated as a high discharge risk critical region.

5. The method as described in claim 1, characterized in that, When generating a collection of Lagrange particles within the water film region of the equivalent water film thickness distribution field, and solving the dynamic equations of the coupled electric field force using smooth particle hydrodynamics to simulate the water film evolution and calculate the probability of conductive water bridge connection, the process includes: Within the water film region of the equivalent water film thickness distribution field, an initial smooth particle set is generated using an adaptive mesh mapping strategy, and the discretization mapping from image pixel data to Lagrange particles is completed through a smooth kernel function. The particle acceleration is obtained by solving the discrete Navier-Stokes equations, and the particle velocity and position are updated by using a predictive-corrected symplectic integrator combined with dynamic smooth length adjustment. The distribution changes of particles on the outer insulation surface at future moments are statistically analyzed to generate a wetting front velocity field. The system detects whether a connected path with a particle density continuously exceeding a preset threshold is formed between the high-voltage end and the grounding end. If such a path exists, the probability of the conductive water bridge being connected is calculated.

6. The method as described in claim 1, characterized in that, The multimodal impedance reconstructor is obtained by training a neural network model based on the Transformer architecture.

7. The method as described in claim 1, characterized in that, The equivalent water film thickness distribution field, the electric field distortion tensor, and the probability of conductive water bridge connection are input into a multimodal impedance reconstructor to generate an electrochemical impedance spectrum. When calculating the dynamic insulation margin attenuation rate based on the electrochemical impedance spectrum, the following steps are included: An initial high-dimensional feature vector is constructed, and principal component analysis is used to reduce the dimensionality of the initial high-dimensional feature vector in the computational domain of the three-dimensional model to obtain a low-dimensional multimodal feature vector. The low-dimensional multimodal eigenvectors are input into the multimodal impedance reconstructor to obtain the electrochemical impedance spectrum; The key resistance parameters were extracted by fitting the electrochemical impedance spectroscopy to the Landers equivalent circuit, and the dynamic insulation margin attenuation rate was calculated by Kalman filtering and differentiation.

8. The method as described in claim 4, characterized in that, When the change in the dynamic insulation margin attenuation rate meets a preset condition or the high discharge risk critical region exceeds a threshold, an early warning is triggered, including: The slope of the dynamic insulation margin attenuation rate curve is monitored, and when the slope exceeds a preset abrupt change threshold, it is determined that the preset condition is met. Monitor the area of ​​the high discharge risk critical region, or calculate the arc induced potential index based on the amplitude of the water film electric field and the maximum eigenvalue of the electric field distortion tensor within the high discharge risk critical region; when the area exceeds a preset area threshold or the arc induced potential index exceeds a preset index threshold, it is determined to exceed the threshold. An early warning signal is generated when any of the judgment conditions are met.

9. The method as described in claim 1, characterized in that, Also includes: The dynamic insulation margin attenuation rate, along with meteorological forecast data and real-time power grid load flow status, are input into a cleaning decision-making agent trained by reinforcement learning to generate a three-dimensional risk heat map and adaptive active intervention instructions. The adaptive active intervention command controls the cleaning equipment or dehumidification device to perform protective actions. The training and decision-making process of the cleaning decision agent includes: using a generative adversarial network to generate a virtual extreme weather evolution scenario based on a historical extreme weather disaster database; driving the cleaning decision agent to perform Monte Carlo tree search in the virtual extreme weather evolution scenario; rehearsing and evaluating the effects of different prevention and control schemes; selecting the optimal prevention and control scheme; and generating the adaptive active intervention instruction based on the optimal prevention and control scheme.

Images