Real-time early warning method for electric shock risk of power equipment

Abstract

The invention discloses a power equipment electric shock risk real-time early warning method, and belongs to the technical field of power safety monitoring. Comprising the following steps: acquiring an ultraviolet image frame containing corona discharge information and a laser radar point cloud frame containing three-dimensional space information; constructing a static probability field fusing the ultraviolet intensity and the local geometrical characteristics of the equipment, wherein the static probability field is used for representing the space probability distribution of a potential corona source; iteratively evolving the static probability field into a dynamic risk prediction field capable of predicting a risk development trend by combining a dynamic evolution equation of an autoregression model and an electric field physical model; and generating an early warning signal based on the dynamic risk prediction field. According to the invention, the physical precursor of electrical breakdown can be sensed, the transformation from passive measurement to active prediction is realized, and the early warning lead and reliability are obviously improved.

Description

Technical Field

[0001] This invention relates to early warning methods, and in particular to a real-time early warning method for the risk of electric shock from power equipment. Background Technology

[0002] With the continuous expansion of the power grid and the sustained increase in voltage levels, the safe and stable operation of power equipment has become crucial. However, in complex operating environments such as ultra-high voltage substations and high-voltage DC converter stations, maintenance personnel inevitably need to have close contact with or operate high-voltage live equipment. These operating environments have extremely high electric field strength and complex and variable hazards. Accidental electric shocks not only pose a fatal threat to the lives of personnel but can also trigger large-scale power grid failures, causing enormous economic losses and social impact.

[0003] Currently, safety monitoring technologies for power operations mainly revolve around computer vision and LiDAR sensing. Visible light-based solutions typically utilize target detection algorithms, such as YOLO or Faster R-CNN, to identify the positions of workers and power equipment from video streams captured by cameras, and estimate safe distances using image coordinates or binocular stereo vision technology. This method effectively utilizes widely deployed surveillance cameras and is relatively low-cost. Another mainstream technology is based on LiDAR, which directly acquires 3D point cloud data of the scene by emitting laser beams and receiving echoes. The backend system denoises, segments, and clusters the point cloud data to identify the 3D contours of personnel and equipment, and directly calculates the Euclidean distance between them in 3D space. Some research has also attempted to combine the two approaches initially; for example, using cameras for target recognition and then roughly defining the areas where LiDAR needs to focus its distance measurement, aiming to comprehensively utilize the advantages of both. These technologies provide certain technical means for safety protection in power operations.

[0004] However, existing technologies have some technical problems when dealing with complex ultra-high voltage power operation scenarios, including problems caused by static risk assessment and oversimplification of risk source models. Summary of the Invention

[0005] The purpose of this invention is to provide a real-time early warning method for the risk of electric shock from power equipment, in order to solve at least one technical problem existing in the prior art.

[0006] Technical solution: A method for real-time early warning of electric shock risk from power equipment, comprising the following steps:

[0007] Acquire and preprocess multimodal data of power equipment operation scenarios to generate time-synchronized ultraviolet image frames and lidar point cloud frames;

[0008] Based on ultraviolet image frames and lidar point cloud frames, a static probability field characterizing the spatial probability distribution of the corona discharge head is constructed.

[0009] By performing time-series evolution of the static probability field through a pre-configured physical model, a dynamic risk prediction field capable of predicting risk development trends is generated.

[0010] Based on the dynamic risk prediction field, the risk level is determined and an early warning signal is generated.

[0011] Traditional risk assessment methods, relying solely on geometric safety distances, suffer from issues such as delayed early warnings and inability to detect invisible ionization processes. By introducing ultraviolet imaging technology sensitive to corona discharge, a proactive risk prediction system can be established, capturing crucial physical precursor signals before electrical breakdown occurs. This represents a shift from traditional passive distance measurement to proactive risk prediction. Specifically, ultraviolet images containing corona discharge information are fused with lidar point clouds containing spatial structure information in a time-synchronized manner to construct a static probability field accurately representing the spatial probability distribution of corona sources. Then, a pre-configured electric field physics model drives temporal evolution, transforming the static probability field into a predictive dynamic risk field. This physics-based evolutionary model can simulate the complete physical process of continuous ionization of air, potentially developing into a transient electric arc, enabling early prediction of the formation and development trends of ionization channels invisible to the naked eye.

[0012] According to one aspect of this application, the time-series evolution of a static probability field is performed using a pre-configured physical model, including:

[0013] Based on the physical model and the historical dynamic risk prediction field of the previous moment, the physical propagation increment characterizing the risk propagation trend in space is calculated.

[0014] By using an autoregressive model, the historical dynamic risk prediction field, the current static probability field, and the physical propagation increment are fused together to update and generate the current dynamic risk prediction field.

[0015] Existing technologies mostly rely on instantaneous state assessment or defining the physical activity range of personnel, lacking the ability to model the temporal evolution of risks. This results in early warning systems only being able to passively respond to dangerous situations that have already occurred. This solution employs a dual-drive mechanism: maintaining historical state memory through an autoregressive model ensures the continuity and stability of risk assessment, avoiding misjudgments caused by instantaneous noise interference; and introducing propagation increment calculations based on an electric field physics model simulates the physical propagation process of risk energy in space. This accurately depicts the dynamic process of corona discharge developing from a localized weak state to a dangerous arc, captures the nonlinear growth characteristics of the risk field, and predicts the development trajectory and growth rate of the risk. Even if the corona source is currently outside a safe distance, if dynamic prediction shows that its energy is propagating at high speed along the electric field lines towards the workers, the system can identify the potential threat in advance.

[0016] According to one aspect of this application, calculating the physical propagation increment characterizing the tendency of risk to propagate in space includes:

[0017] For any neighboring point of any target point in the historical dynamic risk prediction field, determine a direction vector from the neighboring point to the target point, and calculate the alignment between the direction vector and the electric field gradient direction based on the electric field gradient direction of the physical model.

[0018] Based on the field value and alignment of the neighboring point in the historical dynamic risk prediction field, a risk value propagated from the neighboring point to the target point is calculated, and the risk values ​​propagated from all neighboring points are summed to obtain the physical propagation increment of the target point.

[0019] In high-voltage electrical environments, ionization processes preferentially propagate along the direction of maximum electric field gradient, forming potential conductive channels. By calculating the alignment between the direction vector from a neighboring point to the target point and the direction of the electric field gradient at that location, the physical probability of risk energy propagation is quantified. When the two directions are highly aligned, it indicates that the propagation path conforms to the laws of electromagnetic physics, and the probability and intensity of risk propagation increase accordingly.

[0020] Compared to traditional geometric diffusion models, this application can accurately predict the preferred development direction of ionization channels, avoiding the blind assumption of omnidirectional diffusion; by using alignment weighted calculation, the propagation of risk energy strictly follows the electric field distribution law, improving the physical accuracy of dynamic prediction; especially in ultra-high voltage environments with complex equipment layouts, it can accurately identify the most dangerous ionization path.

[0021] According to one aspect of this application, the step of calculating the alignment between the direction vector and the electric field gradient direction specifically includes:

[0022] Calculate the dot product of the direction vector and the electric field gradient direction, and divide it by the product of the magnitudes of the two vectors to obtain the cosine similarity value as the alignment degree.

[0023] A directional alignment assessment was conducted in the field of power safety, specifically calculating the alignment between the physical propagation direction and the electric field gradient direction. The cosine similarity value range is [-1, 1], corresponding to the physical characteristics of electric field propagation. When the cosine value is close to 1, it indicates that the risk propagation direction is highly consistent with the electric field gradient direction, conforming to the physical laws of ionization, and the propagation efficiency is the highest. When the cosine value is close to 0, it indicates that the propagation direction is perpendicular to the electric field, and the propagation efficiency is extremely low. When the cosine value is negative, it indicates that the propagation direction is opposite to the electric field, which does not conform to physical laws. The physical meaning is clear, and the prediction results are interpretable. In practical applications, it can effectively distinguish the degree of danger of different propagation paths.

[0024] According to one aspect of this application, the electric field gradient direction of the physical model is calculated as follows:

[0025] The lidar point cloud frames are segmented to identify the high-voltage conductor point cloud and the grounding point cloud;

[0026] Based on conductor point cloud, grounding point cloud and preset equipment voltage level, the electric field distribution in the space around the high-voltage conductor is estimated, and an electric field gradient vector field containing the electric field gradient direction is generated.

[0027] According to one aspect of this application, the step of estimating the electric field distribution in the space surrounding a high-voltage conductor includes:

[0028] Approximate the conductor point cloud as a simplified geometric primitive of a preset type;

[0029] By combining the grounding point cloud and the equipment voltage level, the equivalent charge distribution corresponding to the simplified geometric primitive is queried from the pre-configured charge density lookup table;

[0030] The electric field gradient vector field is generated by utilizing the equivalent charge distribution.

[0031] Existing methods employ complex numerical methods such as the finite element method for electric field calculations, which, while highly accurate, are computationally time-consuming. This paper proposes a geometric primitive approximation method, simplifying complex conductor surfaces into basic geometric shapes such as line segments and spheres, thus reducing computational complexity and enabling the handling of various complex equipment layouts and voltage levels. LiDAR point cloud segmentation technology identifies high-voltage conductors and grounding equipment, aiming to provide accurate boundary conditions for subsequent electric field calculations.

[0032] First, the complex conductor shape is abstracted into standardized geometric primitives, retaining the key geometric features that affect the electric field distribution and simplifying the calculation model. Then, the equivalent charge distribution under various typical configurations is pre-calculated through offline high-precision simulation, and a multi-dimensional lookup table is constructed to transform the complex calculations at runtime into simple table queries. Finally, the most suitable pre-calculated results are quickly matched according to the geometric parameters and voltage level of the actual scenario.

[0033] According to one aspect of this application, a static probability field characterizing the spatial probability distribution of the corona discharge head is constructed, including:

[0034] Based on ultraviolet image frames and lidar point cloud frames, candidate source point clouds in three-dimensional space are determined;

[0035] For each point in the candidate source point cloud, its corresponding ultraviolet intensity and local geometric features are determined.

[0036] By weighted fusion of ultraviolet intensity and local geometric features, the probability of each point in the candidate source point cloud as a source point is calculated, thereby generating a static probability field.

[0037] Existing technologies identify risk sources as specific points or geometric regions, failing to accurately characterize corona discharge, a physical phenomenon with probabilistic distribution characteristics. In reality, corona discharge is a complex physical process distributed in a probability cloud pattern in space, and its hazard level is not uniformly distributed. This scheme identifies potential risk areas by aggregating candidate source points, avoiding computational redundancy from full-space searches. For each candidate point, it extracts information from two dimensions: ultraviolet intensity and local geometric features. The former serves as direct observational evidence of corona discharge, while the latter serves as indirect evidence of high physical probability. Through a weighted fusion mechanism, it combines direct observational evidence with prior physical knowledge to generate a probability value for each spatial point as a true corona source.

[0038] According to one aspect of this application, a weighted fusion of ultraviolet intensity and local geometric features includes:

[0039] Based on the data quality and distribution of local geometric features of ultraviolet image frames, intensity weight coefficients and geometric weight coefficients are dynamically determined.

[0040] Intensity weighting coefficients are used to weight ultraviolet intensity, and geometric weighting coefficients are used to weight local geometric features;

[0041] Multiply the weighted ultraviolet intensity by the weighted local geometric features to obtain the probability of each point being a source point.

[0042] Specifically, by evaluating the data quality of ultraviolet image frames, the reliability of the current ultraviolet signal is quantified. When the lighting conditions are good and the signal-to-noise ratio is high, the confidence in ultraviolet intensity information is increased. By analyzing the distribution of local geometric features, the distinguishability of geometric information is evaluated. When the surface structure of the device is complex and the geometric features are significant, the dependence on geometric information is increased.

[0043] According to one aspect of this application, dynamically determining the intensity weighting coefficient and the geometric weighting coefficient includes:

[0044] Calculate the signal-to-noise ratio of the ultraviolet signal in the corona spot region of an ultraviolet image frame;

[0045] Calculate the variance of the geometric features of local geometric features;

[0046] Intensity weighting coefficients are determined based on the ultraviolet signal-to-noise ratio, and geometric weighting coefficients are determined based on the geometric characteristic variance.

[0047] The signal-to-noise ratio (SNR) of ultraviolet (UV) signals reflects the reliability of corona discharge observations and is the most direct indicator of the quality of UV intensity information. Geometric feature variance quantifies the complexity of the equipment surface structure and the discriminative power of geometric information; a larger variance indicates more significant geometric features and a greater contribution to risk source localization. By using mapping functions such as the Sigmoid function, these two quality indicators are transformed into normalized weight coefficients, achieving intelligent mapping from data quality assessment to fusion weights. When the UV signal is weak due to distance, the system automatically enhances its focus on geometric features; when the equipment surface is smooth, it relies more on directly observed UV information, significantly improving the system's adaptability.

[0048] According to one aspect of this application, determining local geometric features includes:

[0049] For each point in the candidate source point cloud, determine its neighborhood point set;

[0050] Based on the three-dimensional spatial distribution of the neighborhood point set, the local surface curvature of the point is calculated and used as a local geometric feature.

[0051] According to one aspect of this application, the local surface curvature of a point is calculated based on the three-dimensional spatial distribution of a neighborhood point set, including:

[0052] Principal component analysis is performed on the neighborhood point set to obtain the eigenvalue set;

[0053] Based on the set of eigenvalues, the local surface curvature is obtained through a preset curvature calculation formula.

[0054] Eigenvalues ​​quantify the degree of dispersion of point clouds in different directions; the direction corresponding to the minimum eigenvalue λ0 represents the normal direction of the local surface, and its relative magnitude directly reflects the curvature of the surface; through the normalized curvature formula Curv=λ0 / (λ0+λ1+λ2), the complex geometric features are transformed into dimensionless quantization indices in the range of 0-1.

[0055] Curvature values ​​can accurately identify geometrically abrupt regions such as equipment edges and sharp points, which are physically more prone to corona discharge. They can effectively identify high-risk geometric features such as insulator string connections and conductor branch points, providing reliable physical prior information for probabilistic risk source modeling.

[0056] Beneficial effects include the ability to detect physical precursors of electrical breakdown, enabling a shift from passive measurement to proactive prediction, and significantly improving the lead time and reliability of early warnings. Attached Figure Description

[0057] Figure 1 This is a flowchart of the present invention.

[0058] Figure 2 This is a flowchart of the present invention for the temporal evolution of a static probability field using a pre-configured physical model.

[0059] Figure 3 This is a flowchart of the present invention for calculating the physical propagation increment that characterizes the trend of risk propagation in space.

[0060] Figure 4 This is a flowchart of the electric field gradient direction for calculating the physical model of this invention.

[0061] Figure 5 This is a flowchart illustrating the estimation of the electric field distribution in the space surrounding a high-voltage conductor according to the present invention. Detailed Implementation

[0062] The applicant conducted an in-depth analysis of existing technologies and found that they are generally based on static risk assessment models using geometric safety distances. The core idea of ​​these models is to measure the spatial distance between personnel and equipment and compare it to a fixed safety threshold. The drawback is that they ignore the inherent physical laws governing electrical breakdown accidents, resulting in a passive and delayed risk diagnosis rather than a proactive risk prediction. In ultra-high voltage environments, the real danger does not begin with physical contact between the human body and the conductor, but rather with the ionization and breakdown of the air medium between them. The physical precursors to this process (such as corona discharge) are invisible, indicating the formation of a potential, highly conductive ionization channel. Existing technologies cannot detect these decisive physical precursors and only issue warnings after personnel have entered the dangerous geometric range that could trigger an arc. By this time, the best opportunity for avoidance has often been missed, resulting in insufficient lead time and effectiveness of the warnings.

[0063] Furthermore, existing technologies oversimplify and determinize risk source modeling. Whether identifying a person as a rectangle in a two-dimensional image, a cluster of points in three-dimensional space, or using human posture recognition and prediction to form a motion frame, none of these methods accurately depict the true form of risks under ultra-high voltage environments. Corona discharge itself is a probability cloud with a specific three-dimensional shape and intensity distribution formed at the tip or edge of a conductor; its hazard level is not uniformly distributed in space. Existing technologies cannot effectively model this type of risk source with volumetric and probabilistic characteristics, nor can they accurately assess the non-uniform distribution and dynamic evolution of the risk field in space. In other words, cognitive biases regarding the physical form of risk sources limit the accuracy of risk assessment.

[0064] Therefore, in combination Figures 1 to 5 This invention describes a real-time early warning method for electric shock risk in power equipment, which mainly includes the following steps:

[0065] Acquire ultraviolet image frames containing corona discharge information and lidar point cloud frames containing three-dimensional spatial information;

[0066] A static probability field integrating ultraviolet intensity and local geometric features of the device is constructed to characterize the spatial probability distribution of potential corona sources.

[0067] Furthermore, by combining the dynamic evolution equations of the autoregressive model and the electric field physics model, the static probability field is iteratively evolved into a dynamic risk prediction field that can predict the development trend of risk.

[0068] Finally, an early warning signal is generated based on the dynamic risk prediction field. This invention can sense the physical precursors of electrical breakdown, realizing the transformation from passive measurement to active prediction, and significantly improving the lead time and reliability of the early warning.

[0069] Specifically, the real-time early warning method for the risk of electric shock from electrical equipment includes the following process:

[0070] S1. Acquire the original ultraviolet image stream and the original lidar point cloud stream, align the timestamps through a hardware / software synchronization mechanism, and perform independent preprocessing on the data to obtain synchronized ultraviolet image frames and synchronized lidar point cloud frames that can be used for fusion analysis.

[0071] S11. Acquire the original ultraviolet image stream and the original lidar point cloud stream, and strictly align the timestamps of the two streams (error <1ms) using a hardware trigger signal or Network Time Protocol (NTP) to obtain a timestamp-aligned data stream.

[0072] S12. Read the point cloud portion from the timestamp-aligned data stream, and sequentially perform voxelization downsampling (to reduce computation) and outlier removal (to eliminate noise interference) to obtain a structured and clean synchronous lidar point cloud frame.

[0073] The voxelization downsampling process is as follows:

[0074] 1) Set the voxel mesh size (e.g., 0.05m × 0.05m × 0.05m) based on the point cloud density and computational resource constraints.

[0075] 2) Divide the three-dimensional space into a regular voxel mesh;

[0076] 3) For each non-empty voxel, calculate the centroid coordinates of all points within it;

[0077] 4) Replace all primitive points within the voxel with the centroid;

[0078] 5) Output the downsampled point cloud data.

[0079] S13. Read the image portion from the timestamp-aligned data stream. First, perform Gaussian filtering to smooth the noise. Then, use an adaptive threshold segmentation algorithm to accurately extract the corona discharge region from the background, obtaining the synchronized ultraviolet image frame and the ultraviolet spot binary mask that identifies the spot position.

[0080] S2. Read the synchronous ultraviolet image frame and the synchronous lidar point cloud frame. By combining the light spot intensity and the geometric features of the device with a weighted projection algorithm, the two-dimensional ultraviolet light spot information is mapped into a probability distribution in three-dimensional space to obtain the probability projection field (PPF) that represents the static spatial probability of the corona source at the current moment.

[0081] S21. Input the synchronous lidar point cloud frame and the ultraviolet spot binary mask. Using the pre-calibrated camera intrinsic and extrinsic parameter matrix, construct the view frustum defined by the camera optical center and the ultraviolet spot binary mask contour. Filter out all lidar points falling into the view frustum to obtain the candidate source point cloud.

[0082] S22. Traverse each point in the candidate source point cloud, analyze the spatial distribution of its K-nearest neighbor (KNN) points, calculate the local surface curvature of the point, and obtain a geometric feature set containing the curvature values ​​of each candidate point.

[0083] S23. This step improves upon the limitations of traditional projection methods that rely solely on geometric intersections.

[0084] For each point P in the candidate source point cloud i Perform the following weighted probability calculation:

[0085] Using the camera's intrinsic and extrinsic parameter matrices, the 3D point P is... i The image is then projected back onto the synchronized ultraviolet image frame to obtain its corresponding two-dimensional pixel coordinates (u,v).

[0086] Read the brightness value I(u,v) of the pixel as the intensity factor.

[0087] Query P from the geometric feature set i The corresponding curvature value Curv(P) i ) as a geometric factor.

[0088] P is calculated using the following formula. i The source probability Prob(P) i )=

[0089] ;

[0090] In the formula, the symbols have the following meanings: i is the index of the i-th point in the candidate source point cloud; I(u,v) is the pixel brightness value at coordinates (u,v) in the ultraviolet image frame; w I The weighting coefficient for ultraviolet intensity; w G Here are the weight coefficients for the geometric features; j is the index of the j-th point in the candidate source point cloud, used to traverse all candidate points for normalization; Curv(P i Let P be a point. i The local surface curvature value; Norm() is a normalization function that maps the input value to the [0,1] interval;

[0091] After calculating the probabilities of all candidate points, the data is integrated to form a 3D point cloud data structure in which each point has a probability attribute, which is the final output of this step.

[0092] S3. Read the probability projection field (PPF) at the current moment and the historical PIAF field at the previous moment. Through iterative calculation using the dynamic evolution equation that integrates the autoregressive model and the simplified electric field physical model, obtain the autoregressive evolution field (PIAF) with dynamic physical constraints that can predict the development trend of risk.

[0093] S31. Read the complete point cloud frame of the synchronous lidar. First, identify the high-voltage conductor and the surface of the grounding equipment using a point cloud segmentation algorithm. Then, combined with the known equipment voltage level, use a simplified boundary element method to estimate the direction of the electric field lines in the space surrounding the high-voltage conductor, obtaining the electric field gradient vector field. This field is calculated once during system initialization or updated when the equipment layout changes.

[0094] S32. This step transforms the static probability field into a dynamic and predictable evolutionary field.

[0095] For any point P in the field i Its field value at the current time t is PIAF(P i,t Iterative updates are performed using the following improved autoregressive model:

[0096] Autoregressive component: Inherits the field value PIAF(P) from the previous time step t-1. i , t-1 ), and combined with the current observation value PPF(P i ,t).

[0097] Physical propagation part Δ phys Check P i The nearest point P j If P j The field value PIAF(P) at time t-1 j ,t-1) is higher, and from P j Point to P iThe vector of the electric field gradient vector in P j If the directions are the same, then some energy will be transferred from P. j Spread to P i .

[0098] Final update equation: PIAF(P i ,t)=αPIAF(P i ,t-1)+(1-α)PPF(P i ,t)+Δ phys Where α is the historical forgetting factor, Δ phys For all neighboring points P j The sum of the contributed physical propagation energy. After traversing all relevant points to complete the calculation, the final output is obtained. This output will also serve as the historical PIAF field for the next iteration.

[0099] S4. Analyze the newly generated autoregressive evolution field (PIAF) of physical constraints, quantify the specific risk level by setting a threshold and evaluating the growth rate of energy within the field, and generate the final warning signal based on the level.

[0100] S41. Analyze the autoregressive evolution field (PIAF) of physical constraints, calculate its global energy (the sum of all point field values) and energy growth rate (compared with the global energy at the previous moment), and obtain the risk field quantification index.

[0101] S42. Compare the risk field quantification index with the pre-set multi-level risk thresholds (e.g., energy threshold and growth rate threshold), and determine one of the three levels of safety, caution, or danger based on the range it falls into, thus obtaining the risk level.

[0102] S43. Based on the input risk level, generate the corresponding control command, i.e., the warning signal, and send it to the warning execution module such as the audible and visual alarm.

[0103] In this embodiment, addressing the shortcomings of existing technologies in detecting physical precursors and lacking predictive capabilities, a solution based on a dynamic risk prediction field is proposed. By introducing an ultraviolet camera sensitive to corona discharge, physical signals that are precursors to electrical breakdown are captured in real time. An autoregressive evolution model combining historical states and physical constraints is also constructed. Using the electric field gradient vector field estimated in real time from 3D point cloud data, the propagation direction and growth trend of risk energy in space are deduced. In this way, the formation and development path of invisible ionization channels can be predicted in advance before a fatal electric arc actually occurs, improving the effectiveness and lead time of the early warning.

[0104] To address the oversimplification and determinism in existing risk source modeling techniques, a static probability field-based modeling method is proposed. This method integrates the intensity distribution information of ultraviolet light spots and the local geometric features of laser point clouds (such as surface curvature). Instead of identifying risk sources as fixed points or geometric regions, it calculates the probability value of each candidate point in three-dimensional space as a true source. By dynamically weighting the direct observational evidence of ultraviolet intensity and the indirect evidence of high physical probability of curvature, a three-dimensional probability field accurately characterizing the probability cloud morphology and intensity distribution of corona discharge is constructed. By abandoning simplified geometric assumptions, the basis of risk assessment is closer to physical reality, improving the accuracy of risk source localization and the reliability of subsequent dynamic evolution analysis.

[0105] According to one aspect of this application, S22, the calculation of the geometric features of the candidate source points specifically includes:

[0106] By traversing each point in the candidate source point cloud, and analyzing the spatial distribution of its K-nearest neighbor (KNN) points, the local surface curvature of that point is calculated, resulting in a geometric feature set containing the curvature values ​​of each candidate point.

[0107] S221. For each point P in the candidate source point cloud... i The K-Nearest Neighbors (KNN) search algorithm is used to find the K nearest points (K is a preset parameter, such as K=30) in the dataset where the point is located, and form the neighborhood point set of the point.

[0108] S222. Read the neighborhood point set of each point and calculate the covariance matrix of the point set using Principal Component Analysis (PCA). Perform eigenvalue decomposition on this matrix to obtain three eigenvalues ​​λ0, λ1, and λ2 (λ0 ≤ λ1 ≤ λ2) and their corresponding eigenvectors. The eigenvector corresponding to the smallest eigenvalue λ0 is the eigenvector corresponding to point P. i The local surface normal vector at that location.

[0109] S223. Read the feature value set {λ0,λ1,λ2} for each point, and calculate the dimensionless quantity describing the surface change using the following formula, which is taken as the curvature value Curv(P) for that point. i The larger this value, the closer the area where the point is located is to a linear (such as a cable edge) or a scattered (such as an insulator connection) non-planar structure. Curv(P) i The curvature values ​​of all candidate source points are summed up to form the final geometric feature set.

[0110] According to one aspect of this application, S23, the intensity-geometric weighted probability calculation is specifically as follows:

[0111] S231. To improve the model's adaptability, the weight coefficient w i and w G It is not a fixed value, but is dynamically determined based on the data quality of the current frame.

[0112] By reading the illuminated areas within a synchronized ultraviolet image frame and calculating the ratio of their average brightness to background noise, the signal-to-noise ratio (SNR) of the ultraviolet signal is obtained. UV ).

[0113] Read the complete set of geometric features and calculate the variance Var(Curv) of all curvature values.

[0114] Based on the signal-to-noise ratio and variance, normalized dynamic weight pairs {w} are generated using a preset mapping function (such as the Sigmoid function). i ,w G The logic is: SNR UV The higher, w i The larger the value; the larger Var(Curv) is (indicating high geometric feature discrimination), w G The larger.

[0115] S232. Traverse each point P in the candidate source point cloud. i Perform the following calculations:

[0116] Using the camera intrinsic and extrinsic parameter matrices, P i The three-dimensional coordinates are projected onto the synchronized ultraviolet image frame to obtain the pixel coordinates (u,v) and the brightness I(u,v) is read.

[0117] Query P from the geometric feature set i The corresponding curvature value Curv(P) i ).

[0118] Using the dynamic weights obtained in S231 to pair {w i ,w G} Calculate the unnormalized raw probability Prob at that point. raw (P i )=(w i I(u,v))*(w G Curv(P i ));

[0119] S233. Sum the values ​​from all the original probability sets and calculate the sum Sum(Prob). raw ). Prob for each point raw (P i Dividing by this sum yields the final normalized probability Prob(P). iThe candidate source point cloud is integrated with these one-to-one corresponding normalized probability values ​​to form the final probability projection field (PPF).

[0120] A static probability field is constructed by weighted fusion of ultraviolet spot intensity information and local surface curvature features of laser point clouds. A mechanism that dynamically determines weighting coefficients based on data quality adaptively adjusts the confidence level of the two types of information according to the signal-to-noise ratio of the ultraviolet signal and the salience of geometric features in the current frame. In the complex environment of substations, when the ultraviolet signal is weak due to distance, the system automatically increases its focus on high-curvature areas such as equipment tips and edges (physically more prone to discharge); when the equipment surface is smooth and geometric features are not obvious, it relies more on the directly observed ultraviolet spot intensity. This improves the accuracy and robustness of locating early, weak corona discharge sources. It effectively suppresses noise interference from a single data source and identifies the most potentially dangerous discharge points on complex equipment surfaces, thus providing high-confidence input for subsequent dynamic predictions and significantly reducing the system's false alarm and false negative rates.

[0121] According to one aspect of this application, S31, the estimation of the electric field gradient around the power equipment is specifically as follows:

[0122] S311. Read the complete synchronous lidar point cloud frame and the pre-stored prior model of substation equipment (including the approximate location and type of the equipment). Through registration, use the prior model to guide point cloud segmentation, accurately identify and separate conductor point clouds such as high-voltage conductors and insulator strings, as well as grounding point clouds such as the ground and metal structures.

[0123] S312. This step does not directly solve for the complex charge distribution on the conductor surface, but approximates it as a simple geometric shape.

[0124] The skeleton of a conductor point cloud (such as a cable) is extracted and approximated as a series of live wire segments. Insulators and connecting hardware are approximated as charged spheres.

[0125] Specifically, the approximation of line segments for conductor point clouds includes: 1) using principal component analysis (PCA) to determine the principal direction of the point cloud; 2) projecting the point cloud onto the principal direction and extracting the central axis; 3) segmenting the axis, with each segment having a length of 0.5-2 meters; 4) determining the charge density of each line segment based on its length and cross-sectional radius.

[0126] For the spherical approximation of insulator fittings, the following are included: 1) the center of the bounding box of the point cloud is used as the center of the sphere; 2) the average distance from all points in the point cloud to the center is used as the radius of the sphere; 3) the charge of the equivalent sphere is determined based on its volume and local electric field strength.

[0127] Based on the device voltage level and the position of these simplified geometries relative to the grounding point cloud, the equivalent charge distribution of these simplified geometries is queried and assigned from a charge density lookup table (LUT) that has been pre-calculated through offline accurate simulation.

[0128] Traditional precise electric field simulation algorithms (such as the finite element method) are computationally intensive and cannot meet the response time requirement of less than 200ms for edge-end safety early warning systems. This paper creatively transforms the computationally intensive physical simulation problem into a lightweight geometric matching and lookup table operation by approximating complex conductors identified in lidar point clouds as simplified geometric primitives and using a charge density lookup table (LUT) pre-generated through offline precise simulation to replace the complex real-time solution process. This is the key to enabling the entire dynamic prediction model to be implemented in engineering. It reduces the computational complexity of online estimation of the physical model, making it possible to implement physical model-based risk prediction on resource-constrained edge computing units, resolving the contradiction between complex theoretical algorithms and practical application requirements, and possessing high engineering practical value.

[0129] S313. Within the spatial region requiring risk assessment (e.g., the area covered by the probabilistic projection field (PPF), generate a three-dimensional mesh. For each node in the mesh, according to Coulomb's law, superimpose the electric field vectors generated by all charge elements in the equivalent charge distribution to calculate the electric field intensity and direction of that node. This ultimately forms a discretized electric field gradient vector field.

[0130] According to one aspect of this application, S32, the fusion evolution of autoregression and physical propagation, specifically:

[0131] S321. For each point P in the field... i First, the observations at the current moment and historical memories are integrated. The probability projection field (PPF) at the current moment and the historical PIAF field at the previous moment are read, and the intermediate evolution field (P) without physical propagation is calculated using the following formula. i,t )=α*PIAF(P i , t-1 )+(1-α)PPF(P i ,t); where α is the preset historical forgetting factor (e.g., 0.7).

[0132] S322. For each point P in the intermediate evolution field i Calculate its physical propagation increment Δ phys (P i Specifically:

[0133] Initialize Δ phys (P i )=0.

[0134] Find P i The neighbor set {P} in the intermediate evolution field j}

[0135] For each neighboring point P j Calculate from P j Point to P i Direction vector V j i.

[0136] Queries point P from the electric field gradient vector field. j electric field gradient direction at E(P j ).

[0137] Calculate the alignment (cosine similarity) of two vectors: Alignment = cos(V) ji ,∇ E(P j )).

[0138] If Alignment is positive (indicating that the propagation direction is the same as the electric field direction), then the calculation from P... j Spread to P i Energy Transfer = β * IntermediateField(P) j,t )*Alignment, where β is the propagation rate coefficient.

[0139] Add up the energy contributed by all neighbors: Δ phys (P i =Σ(Transfer). By summing the increments of all points, we obtain the complete physical propagation increment field.

[0140] The formula for calculating the direction vector V_ji is: V_ji=(P_i-P_j) / ||P_i-P_j||;

[0141] P_i=(x_i,y_i,z_i) represents the three-dimensional coordinates of the target point; P_j=(x_j,y_j,z_j) represents the three-dimensional coordinates of the neighboring points; P_i-P_j represents the vector difference between the two points; ||P_i-P_j|| represents the Euclidean norm (modulus) of the vector; V_ji represents the unit direction vector from point P_j to point P_i.

[0142] S323. Superimpose the intermediate evolution field with the physical propagation increment field to obtain the final field state at the current time t. PIAF(P i ,t)=IntermediateField(P i ,t)+Δ phys (P iThis result is the final output of this step, and it will be cached as the input of the historical PIAF field for the next time step.

[0143] This step introduces an autoregressive model combining historical states and physical propagation increments based on electric field gradients to construct a dynamic risk prediction field capable of temporal evolution. This enables the method to simulate and extrapolate the entire dynamic development process of risk. In specific scenarios of ultra-high voltage power operations, it can accurately simulate the complete physical process of continuous ionization of the air medium, potentially developing into a transient electric arc, based on the currently formed weak corona discharge (characterized by a static probability field) and the electric field line direction estimated from the equipment's three-dimensional model. This represents a shift from passive ranging to active prediction, enabling the early prediction of the formation and development trend of ionization channels invisible to the naked eye, thereby improving the lead time for early warnings and the ability to detect sudden discharge accidents, and enhancing the safety of operators.

[0144] According to another aspect of this application, a real-time early warning method for electric shock risk of power equipment is provided. This method can be applied to an embedded early warning system or a central server. Its core processing flow includes data acquisition and preprocessing, static probability field construction, dynamic risk prediction field generation, and risk judgment and early warning.

[0145] Step 1: Acquire and preprocess multimodal data.

[0146] In this embodiment, a multimodal data acquisition module deployed in the power equipment operation scenario is used to simultaneously acquire an ultraviolet image stream containing corona discharge information and a lidar point cloud stream containing three-dimensional spatial information.

[0147] The multimodal data acquisition module may include a solar-blind ultraviolet camera sensitive to wavelengths in the 240-280nm range and a lidar with a scanning frequency of not less than 10Hz. Hardware timestamp synchronization technology ensures that the time error between the two data streams is less than 1ms.

[0148] Furthermore, the acquired data stream is preprocessed, for example, by voxel downsampling the lidar point cloud stream (e.g., setting the voxel size to 0.1m). 3 The ultraviolet image stream is subjected to Gaussian filtering and adaptive thresholding after outlier removal to obtain clear and aligned LiDAR point cloud frames and ultraviolet image frames required for subsequent steps.

[0149] Step 2 involves constructing a static probability field characterizing the spatial probability distribution of the corona discharge source. This step aims to accurately map two-dimensional ultraviolet spot information to a probabilistic distribution of risk sources in three-dimensional space. Specifically, this may include:

[0150] Based on the pre-calibrated spatial pose relationship (i.e., camera intrinsic and extrinsic parameter matrix) between the ultraviolet camera and the lidar, the spot area in the ultraviolet image frame is projected in reverse into three-dimensional space to form an observation cone. All points that fall within this cone and originate from the lidar point cloud frame constitute a candidate source point cloud.

[0151] For each point in the candidate source point cloud, extract its features in two dimensions:

[0152] First, its three-dimensional coordinates are projected back into the ultraviolet image frame, and the brightness value of the corresponding pixel is read as its ultraviolet intensity.

[0153] Secondly, calculate its local geometric features.

[0154] In this embodiment, the local geometric feature is the local surface curvature. First, for any point, its neighborhood point set is determined using the K-nearest neighbor algorithm (e.g., K=30). Then, principal component analysis (PCA) is performed on this neighborhood point set to obtain three eigenvalue sets {λ0, λ1, λ2}, and the local surface curvature of that point is calculated based on the formula Curv=λ0 / (λ0+λ1+λ2). A brief description of the KD-tree K-nearest neighbor search process is as follows:

[0155] Basic implementation steps of KD-tree K-nearest neighbor search:

[0156] 1) Constructing a KD-tree: Recursively divide the points according to different dimensions to construct a binary tree structure;

[0157] 2) Search process:

[0158] a) Starting from the root node, recursively descend to the leaf node according to the coordinates of the query point;

[0159] b) Select the leaf node as the current nearest neighbor candidate;

[0160] c) During the backtracking process, check whether another subtree of each node might contain a closer point;

[0161] d) Maintain a priority queue of size K to store the K nearest neighbors found so far;

[0162] 3) Return the K points in the priority queue as the neighborhood point set.

[0163] Features are weighted and fused to generate a static probability field: First, the weight coefficients are dynamically determined. For example, the signal-to-noise ratio of the ultraviolet signal is obtained by calculating the signal-to-noise ratio of the spot region in the ultraviolet image frame, and the variance of the geometric features is obtained by calculating the variance of the curvature of all local surfaces.

[0164] Based on these two indicators, an intensity weighting coefficient and a geometric weighting coefficient are determined through a preset mapping function. Preferably, when the ultraviolet signal is clear (high signal-to-noise ratio), the intensity weighting coefficient is increased; when the device geometry is complex (large curvature variance), the geometric weighting coefficient is increased.

[0165] Subsequently, the ultraviolet intensity and local surface curvature of each candidate point are multiplied by their corresponding weighting coefficients, and then the two are multiplied together to obtain the probability that the point is the source point. Finally, the probabilities of all candidate points are normalized to obtain the final static probability field.

[0166] Through the above steps, this method no longer treats the risk source as a simple three-dimensional point, but generates a probabilistic risk distribution map that more realistically reflects the physical reality. This provides high-precision and high-confidence input for subsequent dynamic prediction and reduces misjudgments caused by inaccurate source location.

[0167] Step 3: Generate a dynamic risk prediction field through temporal evolution. The aim is to evolve a static risk probability distribution into a prediction field with both temporal and physical dimensions. Specifically, this may include:

[0168] Estimating the electric field gradient vector field. In this embodiment, the complete lidar point cloud frame is first semantically segmented to identify conductor point clouds such as insulator strings and high-voltage bushings, as well as grounding point clouds such as towers and the ground. Then, the identified conductor point clouds are approximated as simplified geometric primitives such as line segments and spheres. Based on the positions of these simplified geometric primitives relative to the grounding point clouds and the preset equipment voltage level (e.g., 500kV), an equivalent charge distribution is assigned to these primitives from a pre-generated charge density lookup table (LUT) generated through offline electromagnetic simulation. Finally, based on this equivalent charge distribution, the electric field distribution within the working space is calculated using Coulomb's law, thereby obtaining the electric field gradient vector field.

[0169] Specifically, the current dynamic risk prediction field is jointly determined by the historical dynamic risk prediction field from the previous moment and the current static probability field. First, a physical propagation increment is calculated: for any target point within the field, its neighboring points are analyzed. If the direction vector from a neighboring point to the target point has a high degree of alignment with the electric field gradient direction at that neighboring point (preferably quantified by calculating the cosine similarity between the two), then a portion of the field value (risk energy) from the neighboring point is propagated to the target point proportionally according to the alignment. Then, through an autoregressive model, the historical dynamic risk prediction field, the current static probability field, and the calculated physical propagation increment are fused to update and generate the current dynamic risk prediction field.

[0170] Through the steps described above, this method elevates risk assessment from a static to a dynamic level. For example, even if a corona source is currently outside a safe distance, the system can still detect the risk in advance if the dynamic risk prediction field shows that its energy is rapidly propagating and increasing along the electric field lines towards the workers. Tests show that, compared to the traditional fixed-distance threshold method, this method can increase the effective warning lead time by approximately 300-500 ms.

[0171] Step 4: Based on the dynamic risk prediction field, determine the risk level to generate an early warning signal.

[0172] In this embodiment, multiple indicators, such as the total global energy value, energy growth rate, and maximum extension boundary of the newly generated dynamic risk prediction field, are analyzed and compared with a multi-level risk threshold model. For example, when the total energy value of the field exceeds threshold T1 or the energy growth rate exceeds T2, it is determined to be a medium-level risk; when the predicted extension boundary of the field will reach the safe operating radius within 200ms, it is determined to be a high-level risk. Finally, based on the determined risk level, an early warning signal is generated and the early warning execution modules, such as audible and visual alarms, are activated.

[0173] Through the above complete technical solutions, the present invention works together to generate a brand-new risk warning capability based on physical process prediction, realizing a leap from spatial distance measurement to temporal trend prediction, and improving the reliability and foresight of personnel safety protection in ultra-high voltage power operation scenarios.

[0174] Example 2: This example details the real-time estimation process of the electric field gradient vector field.

[0175] In this embodiment, the lidar point cloud frames are segmented. Specifically, firstly, the RANSAC (Random Sample Consensus) algorithm is used to fit the ground plane in the scene with high accuracy and mark it as the ground point cloud. Subsequently, in the remaining point cloud after filtering out the ground points, registration is performed using a prior model of the substation equipment (e.g., a simplified CAD model containing the approximate location and size of the equipment), and clustering is performed using the DBSCAN (Density-Based Spatial Clustering) algorithm. Preferably, geometric shape matching is performed on each cluster; for example, high-voltage bushings and conductors are identified by fitting a cylindrical model, and equipment surfaces such as switchgear are identified by fitting a planar model, thereby accurately separating the conductor point cloud.

[0176] The conductor point cloud obtained in the previous step is geometrically approximated. For example, for a point cloud identified as a conductor, it is approximated by a series of simplified geometric primitives connected end to end through its central axis—a live wire segment; for a point cloud identified as an insulator fitting or equipment connection, it is approximated by a simplified geometric primitive—a charged sphere.

[0177] Furthermore, based on the key dimensions of these simplified geometric primitives (such as line segment length and sphere radius), their average height relative to the grounding point cloud, and the preset equipment voltage level, the system queries a pre-configured charge density lookup table for the corresponding equivalent charge distribution. It should be noted that the charge density lookup table is a multi-dimensional table whose index includes geometric primitive type, key dimensions, and height relative to ground. The values ​​stored in the table are equivalent charge density coefficients pre-calculated using offline high-precision electromagnetic simulation software (such as COMSOL). For example, for a line segment primitive with a length of 2 meters and an average height of 15 meters above the ground at a 500kV voltage level, the system can obtain its equivalent linear charge density as -1.2 μC / m by looking up the table.

[0178] Through the above steps, this embodiment transforms the complex electromagnetic field boundary value problem into a geometric fitting and table lookup operation that can be completed in milliseconds, providing key technical support for achieving the real-time performance of the entire early warning method, and ultimately generating the electric field gradient vector field necessary for dynamic evolution.

[0179] Example 3: Implementation of static probability field construction:

[0180] In this embodiment, the ultraviolet signal signal-to-noise ratio (SNR) UV The average pixel value I of the spot region in the ultraviolet image frame can be calculated. signal Standard deviation σ of pixel values ​​in the background region adjacent to the spot noise The ratio is obtained. The geometric feature variance Var(Curv) is obtained by directly statistically analyzing the variance of the local surface curvature values ​​of all points in the candidate source point cloud. Preferably, the intensity weighting coefficient w i and geometric weighting coefficient w G Normalization and mapping can be performed using the Sigmoid function, for example: w i =1 / (1+exp(-k1*(SNR UV -T1)))w G =1 / (1+exp(-k2*(Var(Curv)-T2))) where k1,k2,T1,T2 are preset parameters calibrated based on experimental data, used to control the sensitivity and offset of the weights.

[0181] For any point P in the candidate source point cloud, its ultraviolet intensity and local geometric features (i.e., curvature) are first normalized. Then, the intensity weighting coefficient w obtained in the previous step is used. i and geometric weighting coefficient w G The values ​​are weighted, and the two weighted feature values ​​are multiplied together to obtain the unnormalized probability value of that point. For example, in a certain frame, w is calculated. i =0.8, w G=0.3. For candidate point P, its normalized ultraviolet intensity is 0.9, and its normalized curvature is 0.6, so its unnormalized probability is (0.8*0.9)*(0.3*0.6)=0.1296. Finally, the unnormalized probabilities of all candidate points are globally normalized (so that their sum is 1), thus generating the final static probability field.

[0182] Through this adaptive dynamic weighting method, this embodiment enables the construction process of the static probability field to intelligently respond to changes in data quality under different scenarios. When the quality of one information source deteriorates, it can automatically enhance the dependence on another reliable information source, thereby ensuring the continuous high accuracy and high reliability of risk source location.

[0183] Example 4: Implementation of local geometric feature calculation:

[0184] In this embodiment, to improve search efficiency, the lidar point cloud frame can be pre-constructed as a k-dimensional tree (KD-tree) data structure. For any point P in the candidate source point cloud set, its neighborhood point set can be efficiently obtained by performing a K-nearest neighbor search in the KD-tree (for example, the value of K can be adaptively adjusted according to the point cloud density, preferably in the range of 20-50).

[0185] After obtaining the neighborhood point set of a given point, principal component analysis is performed on it. Specifically, firstly, the 3x3 covariance matrix of the point set relative to its centroid is calculated. Then, eigenvalue decomposition is performed on the covariance matrix to obtain three eigenvalue sets {λ0, λ1, λ2} arranged in ascending order. It should be noted that the eigenvalues ​​characterize the dispersion of the point set in three mutually orthogonal principal directions. Among them, the direction corresponding to the smallest eigenvalue λ0 is approximately the surface normal direction at that point. In this embodiment, the formula Curv=λ0 / (λ0+λ1+λ2) is preferably used to calculate the local surface curvature. The physical meaning of this formula is that when the point set is ideally planar, λ0 approaches 0, and the curvature also approaches 0; while when the point set is linear (such as the edge of a device) or point-like (such as the corner of a device) distribution, the proportion of λ0 relative to the other two eigenvalues ​​increases, and the curvature value also increases accordingly, thereby effectively distinguishing between planar and non-planar regions.

[0186] By using this curvature quantization method based on principal component analysis, this embodiment can stably and accurately extract numerical features from point clouds that characterize geometric abrupt changes in regions such as device edges and tips, providing a reliable quantitative basis for identifying high-risk points (i.e., weighted fusion processes) that are physically more prone to corona discharge.

[0187] The preferred embodiments of the present invention have been described in detail above. However, the present invention is not limited to the specific details in the above embodiments. Within the scope of 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 fall within the protection scope of the present invention.

Claims

1. A method for real-time early warning of electric shock risk from power equipment, characterized in that, include: Acquire and preprocess multimodal data of power equipment operation scenarios to generate time-synchronized ultraviolet image frames and lidar point cloud frames; Based on ultraviolet image frames and lidar point cloud frames, a static probability field characterizing the spatial probability distribution of the corona discharge head is constructed. By performing time-series evolution of the static probability field through a pre-configured physical model, a dynamic risk prediction field capable of predicting risk development trends is generated. Based on the dynamic risk prediction field, the risk level is determined and an early warning signal is generated.

2. The method according to claim 1, characterized in that, The static probability field is subjected to temporal evolution using a pre-configured physical model, including: Based on the physical model and the historical dynamic risk prediction field of the previous moment, the physical propagation increment characterizing the risk propagation trend in space is calculated. By using an autoregressive model, the historical dynamic risk prediction field, the current static probability field, and the physical propagation increment are fused together to update and generate the current dynamic risk prediction field.

3. The method according to claim 2, characterized in that, Calculate the physical propagation increment that characterizes the trend of risk propagation in space, including: For any neighboring point of any target point in the historical dynamic risk prediction field, determine a direction vector from the neighboring point to the target point, and calculate the alignment between the direction vector and the electric field gradient direction based on the electric field gradient direction of the physical model. Based on the field value and alignment of the neighboring point in the historical dynamic risk prediction field, a risk value propagated from the neighboring point to the target point is calculated, and the risk values ​​propagated from all neighboring points are summed to obtain the physical propagation increment of the target point.

4. The method according to claim 3, characterized in that, The alignment between the direction vector and the electric field gradient direction is calculated as follows: Calculate the dot product of the direction vector and the electric field gradient direction, and divide it by the product of the magnitudes of the two vectors to obtain the cosine similarity value as the alignment degree.

5. The method according to claim 3, characterized in that, The electric field gradient direction of the physical model is calculated as follows: The lidar point cloud frames are segmented to identify the high-voltage conductor point cloud and the grounding point cloud; Based on conductor point cloud, grounding point cloud and preset equipment voltage level, the electric field distribution in the space around the high-voltage conductor is estimated, and an electric field gradient vector field containing the electric field gradient direction is generated.

6. The method according to claim 5, characterized in that, Estimating the electric field distribution in the space surrounding a high-voltage conductor includes: Approximate the conductor point cloud as a simplified geometric primitive of a preset type; By combining the grounding point cloud and the equipment voltage level, the equivalent charge distribution corresponding to the simplified geometric primitive is queried from the pre-configured charge density lookup table; The electric field gradient vector field is generated by utilizing the equivalent charge distribution.

7. The method according to claim 1, characterized in that, Constructing a static probability field characterizing the spatial probability distribution of the corona discharge head, including: Based on ultraviolet image frames and lidar point cloud frames, candidate source point clouds in three-dimensional space are determined; For each point in the candidate source point cloud, its corresponding ultraviolet intensity and local geometric features are determined. The ultraviolet intensity and local geometric features are weighted and fused to calculate the probability of each point in the candidate source point cloud as a source point, thereby generating a static probability field.

8. The method according to claim 7, characterized in that, Weighted fusion of ultraviolet intensity and local geometric features includes: Based on the data quality and distribution of local geometric features of ultraviolet image frames, intensity weight coefficients and geometric weight coefficients are dynamically determined. Intensity weighting coefficients are used to weight ultraviolet intensity, and geometric weighting coefficients are used to weight local geometric features; Multiply the weighted ultraviolet intensity by the weighted local geometric features to obtain the probability of each point being a source point.

9. The method according to claim 8, characterized in that, Dynamically determine the intensity weighting coefficient and the geometric weighting coefficient, including: Calculate the signal-to-noise ratio of the ultraviolet signal in the corona spot region of an ultraviolet image frame; Calculate the variance of the geometric features of local geometric features; Intensity weighting coefficients are determined based on the ultraviolet signal-to-noise ratio, and geometric weighting coefficients are determined based on the geometric characteristic variance.

10. The method according to claim 7, characterized in that, Determine local geometric features, including: For each point in the candidate source point cloud, determine its neighborhood point set; Based on the three-dimensional spatial distribution of the neighborhood point set, the local surface curvature of the point is calculated and used as a local geometric feature; Among them, the local surface curvature of a point is calculated based on the three-dimensional spatial distribution of the neighborhood point set, including: Principal component analysis is performed on the neighborhood point set to obtain the eigenvalue set; Based on the set of eigenvalues, the local surface curvature is obtained through a preset curvature calculation formula.

Images