A Dynamic Modeling Method for Safe Zones of Live Bodies in Substations Based on Electric Field-Point Cloud Coupling Compensation

Abstract

This invention discloses a dynamic modeling method for the safety zone of a energized body in a substation based on electric field-point cloud coupling compensation, belonging to the field of substation safety detection technology. The method involves: acquiring original point cloud data and spatial electric field gradient data of the energized body; compensating the original point cloud data using an electric field distortion compensation method; constructing a deformation CNN point cloud reconstruction model to reconstruct the original point cloud data and spatial electric field gradient data; and delineating the safety zone layer by layer based on the reconstructed point cloud coordinates and real-time electric field intensity. This invention effectively reduces distortion during point cloud data acquisition through the electric field distortion compensation method, improving the accuracy and reliability of the point cloud data; accurately reconstructing the three-dimensional structure of the energized body using a deformation CNN point cloud reconstruction model; and more accurately delineating the safety boundary of the energized area through the combination of real-time electric field intensity data and point cloud data, reducing false detections and missed detections, and improving the timeliness and accuracy of safety warnings.

Description

Technical Field

[0001] This invention belongs to the field of substation safety detection technology, specifically involving a dynamic modeling method for the safety area of ​​energized bodies in substations based on electric field-point cloud coupling compensation. Background Technology

[0002] As a crucial component of the power system, the safe operation of substations directly impacts the stability of the power grid and the safety of personnel. However, substations present complex environments characterized by high voltage, high risk, and dense equipment, making them prone to accidents. For example, personnel accidentally entering dangerous areas, equipment malfunctions, and foreign object intrusion can all lead to serious consequences.

[0003] Currently, safety monitoring of energized areas in substations mainly relies on image recognition and point cloud data processing technologies. Regarding point cloud data processing, existing research has developed deformation monitoring methods using multi-temporal lidar point cloud data with tools such as MATLAB. These methods project the point cloud onto a deformation field grid and use the M3C2 algorithm to calculate the deformation, thereby achieving dynamic monitoring of structural deformation.

[0004] In terms of electric field compensation, although existing research has involved the acquisition and processing of electric field information, how to couple electric field information with point cloud data to improve the accuracy of charged region modeling remains an unresolved problem.

[0005] Chinese patent CN118351672B discloses an intelligent detection and early warning method and system for energized areas in substations. The method includes: identifying equipment entities in a target area image to obtain multiple equipment entity wireframe information and multiple wireframe accuracy information; performing iterative point cloud fitting to obtain multiple equipment entity point clouds and multiple fitting accuracy information; constructing the energized area range within a three-dimensional model of the area based on the electric field information of multiple equipment entities; and measuring the distance to personnel to generate a safety warning level for early warning. This application addresses the technical problem of poor accuracy and comprehensiveness in energized area detection and early warning due to poor accuracy in energized area setting and personnel distance measurement, leading to frequent false detections and missed detections. It improves the accuracy and comprehensiveness of energized area detection and early warning, thereby ensuring the safe and stable operation of substations.

[0006] The patent application for an intelligent detection and early warning method and system for energized areas of substations has the following shortcomings: when processing point cloud data, the influence of electric field distortion on the point cloud data is often ignored, resulting in a deviation between the reconstructed three-dimensional model and the actual energized area, which in turn affects the accuracy of safety early warning.

[0007] Therefore, the technical problem that this invention aims to solve is how to resolve the data deviation problem when effectively coupling electric field information with point cloud data, thereby improving the accuracy of safety warnings. Summary of the Invention

[0008] The purpose of this invention is to provide a dynamic modeling method for the safe area of ​​energized bodies in substations based on electric field-point cloud coupling compensation, so as to solve the problems mentioned in the background art.

[0009] The objective of this invention is achieved as follows: a dynamic modeling method for the safe zone of a energized body in a substation based on electric field-point cloud coupling compensation, characterized by the following steps:

[0010] Step S1: Collect the original point cloud data and spatial electric field gradient data of the charged body;

[0011] Step S2: Compensate the original point cloud data of the charged body using the electric field distortion compensation method;

[0012] Step S3: Construct a deformation CNN point cloud reconstruction model to reconstruct the original point cloud data and spatial electric field gradient data;

[0013] Step S4: Delineate safe zones based on the reconstructed point cloud coordinates and real-time electric field intensity.

[0014] Preferably, the acquisition of the original point cloud data and spatial electric field gradient data of the charged body in step S1 specifically includes:

[0015] Step S1-1: Synchronously acquire raw point cloud data of the charged body using a rigidly connected laser scanner, a broadband electric field sensor, and an IMU attitude compensation unit. and spatial electric field gradient data ;

[0016] Step S1-2: Synchronization data acquisition triggered by GPS-disciplined clock source:

[0017] Laser scanner outputs raw point cloud data The electric field sensor collects spatial electric field gradient data as follows:

[0018] ;

[0019] in, for Gradient along the axis; for Gradient along the axis; for Gradient along the axis;

[0020] Step S1-3: Convert the Euler angles output by the IMU into a rotation matrix using a coordinate transformation matrix. ,make sure Angle with the point cloud normal vector It satisfies the formula: .

[0021] Preferably, in step S2, an electric field distortion compensation method is used to compensate the original point cloud data, specifically as follows:

[0022] S2-1: Classify insulator types based on point cloud RGB features, and retrieve dielectric constant compensation coefficients from the compensation coefficient table stored in the FPGA. ;

[0023] Dielectric constant compensation coefficient Using a dielectric constant compensation coefficient table stored in the FPGA module, material type is matched in real time; the compensation calculation cycle is ≤10 ms, and the delay error is <0.1%;

[0024] when At that time, porcelain insulators were activated; When using silicone rubber bushings; porcelain insulators HSV color gamut: H180-240°, S>0.2, V 30%-60%; silicone rubber bushings: H 30-60°, S<0.5, V 50%-80%;

[0025] S2-2: Utilizing raw point cloud data Real-time electric field distortion compensation calculation is performed to generate compensated original point cloud data. :

[0026] ;

[0027] in, The raw point cloud data, The sampling interval; Let be the electric field intensity gradient of the i-th element; This represents the number of insulators.

[0028] Preferably, the deformation CNN point cloud reconstruction model adopts a 7-layer network architecture, including an input layer, a first execution layer, a second execution layer, a third execution layer, a fourth execution layer, a fifth execution layer, and an output layer;

[0029] The first execution layer includes a feature extraction module and a basic deformation prediction module; the second execution layer includes a downsampling module and a gradient prediction module; the third execution layer includes a dilated convolution module and a Jacobi calculation module; the fourth execution layer includes a convolution enhancement module, a curvature compensation module, a convolution compression module and an attention module; and the fifth execution layer includes an upsampling module and a deformation synthesis module.

[0030] The feature extraction module uses an 11×11×11 convolution kernel to extract spatial features; the basic deformation prediction module uses a 3×3×3 deformation convolution kernel; the downsampling module uses a 2×2×2 three-dimensional max pooling; the gradient prediction module uses a 3×3×3 deformation convolution to calculate the offset gradient; the dilated convolution module uses a three-dimensional dilated convolution; and the Jacobian calculation module uses the Jacobian to calculate the Hessian operator.

[0031] The convolution enhancement module uses 3×3×3 deformable convolution, the curvature compensation module calculates the curvature tensor; the convolution compression module uses 3×3×3 convolution, and the attention module is used to calculate the attention to the component space.

[0032] The upsampling module performs 4x upsampling, and the deformation synthesis module uses 3×3×3 deformation convolution.

[0033] Preferably, in step S3, constructing a deformation CNN point cloud reconstruction model to reconstruct the original point cloud data and spatial electric field gradient data of the charged body specifically involves:

[0034] Step S3-1: Input spatial coordinates and electric field gradient using the input layer:

[0035] Input tensors to construct compensated original point cloud data With electric field gradient Concatenate into a 6-channel input feature tensor ;

[0036] Step S3-2: Implement basic deformation awareness from the input layer to the first execution layer:

[0037] The feature extraction module performs the DeformConv1 operation, extracting spatial features through an 11×11×11 convolution kernel. The basic deformation prediction module performs the DeformConv2 operation, predicting the basic offset using a 3×3×3 deformation convolution kernel. Output: , ;

[0038] in, The number of points in the point cloud; These are the weights of the 3D convolution kernel. This is a 3D convolution operation; It is the bias vector; It is a linear rectification activation function; This is the first stage feature map; Predict the convolution kernel for deformation; Use the Sigmoid activation function; Base offset vector;

[0039] Step S3-3: Refine the deformation field from the first execution layer to the second execution layer:

[0040] The downsampling module performs MaxPool operations. The gradient prediction module performs a 2×2×2 three-dimensional max pooling operation and a DeformConv3 operation, using a 3×3×3 deformable convolution to calculate the offset gradient. Output: , ;

[0041] in, For three-dimensional max pooling, Features after pooling; These are the kernel weights; The activation function is set to a leakage coefficient of 0.01. This is the second-stage feature map; Predict the convolution kernel for gradient; For the offset gradient field;

[0042] Step S3-4: Implement curvature-sensitive processing from the third execution layer to the fourth execution layer:

[0043] The dilated convolution module performs a Conv4 (d=2) operation, a dilated convolution with a dilation rate of d=2, to expand the receptive field. The Jacobi calculation module is based on the displacement gradient field. Compute the Hessian operator:

[0044] ;

[0045] pass Adding a penalty term to the network loss function To prevent point cloud from being partially folded or overstretched, output , ;

[0046] in: 3D dilated convolution with dilation rate 2; This is the third-stage feature map; For Jacobian matrix fields;

[0047] Step S3-5: Implement curvature compensation from the third execution layer to the fourth execution layer:

[0048] Using the DeformConv5 convolution enhancement module The DeformConv5 operation is a 3×3×3 deformable convolution to enhance features;

[0049] The curvature tensor is calculated using the curvature compensation module, and the surface curvature is calculated using the second-order partial derivative. Output , ;

[0050] in, These are the kernel weights; This is the fourth stage feature map; For curvature tensor; The magnitude of the electric field gradient; For curvature compensation amount, This represents the geometric fit coefficient, used to compensate for normal distortion caused by the electric field gradient; It is a 1×1 convolution (dimension adjustment); This is a residual feature map;

[0051] Step S3-6: Implement field suppression from the fourth execution layer to the fifth execution layer:

[0052] Perform Conv6 operations using the convolutional compression module Conv6 uses 3×3×3 convolutions, which compress and reduce the dimensionality of features;

[0053] Calculate the attention to the component space using the attention module based on the electric field method. , and Threshold binding embeds safety procedure thresholds into neural network parameters, as follows: Output ;

[0054] in, It is a 3D convolution; This is the feature map for the fifth stage; This refers to the normal component of the electric field; Spatial attention weights; This is element-wise multiplication; For weighted feature maps;

[0055] Step S3-7: Deformation synthesis is achieved from the fifth execution layer to the output layer:

[0056] Perform 4x upsampling using the upsampling module, followed by trilinear interpolation. The original number of points is restored; the deformation synthesis module performs the DeformConv7 operation, which is a 3×3×3 deformation convolution, to generate the final displacement. Output ;

[0057] in, This is a 4x trilinear interpolation upsampling; For upsampling features; To output the convolution kernel, It is the hyperbolic tangent activation function; This represents the final deformation displacement;

[0058] Step S3-8: Enhance edge details along the electric field normal ;

[0059] in, This is the edge enhancement coefficient;

[0060] Step S3-9: Final Output Synthesis .

[0061] Preferably, the deformation CNN point cloud reconstruction model includes a master device and a slave device, with the slave device employing a TensorR module;

[0062] Edge deployment optimization is achieved using a deformation CNN point cloud reconstruction model, specifically as follows:

[0063] First, the convolutional network of the deformation CNN point cloud reconstruction model is dynamically divided into three layers based on computational cost:

[0064] Lightweight: DeformConv1-DeformConv3, computational cost TOPS;

[0065] Medium-scale: DeformConv4, computational cost TOPS;

[0066] Heavyweight: DeformConv5-DeformConv7, computational complexity TOPS;

[0067] Secondly, lightweight and medium-weight tiers are assigned to the slave device TensorR module, and heavyweight tiers are assigned to the master device to meet resource constraints. in Indicates equipment Maximum computing power;

[0068] Real-time resource monitoring, collecting the following device status:

[0069] Equipment utilization rate: ;

[0070] Load change rate: ;in, Second, This represents the utilization rate of device k at time t;

[0071] Finally, implement the dynamic degradation strategy:

[0072] When minimum resource margin TOPS:

[0073] Adjust rendering resolution: Computational scaling: scaling factor Choose according to the following rules:

[0074] .

[0075] Preferably, the deformation CNN point cloud reconstruction model has migration triggering conditions in achieving edge deployment optimization, specifically:

[0076] When the device The migration mechanism is triggered when any of the following conditions are met:

[0077] And duration Second; $U_k(t) > 0.95;

[0078] The migration process includes the following three stages:

[0079] Freeze: Saves the layer group state, time consumed. ms;

[0080] Transmission: The compressed status data is sent via the TSN network, taking [time]. ms;

[0081] Restore: Target device loading state, time taken ms.

[0082] Preferably, in step S4, the safe zone is delineated in layers based on the reconstructed point cloud coordinates and real-time electric field intensity, specifically as follows:

[0083] Step S4-1: Based on reconstructed point cloud data Constructing implicit surfaces of charged bodies Calculate any point Minimum Euclidean distance to the implicit surface of a charged body:

[0084] :

[0085] Step S4-2: When the current change of the charged body reaches When the safety boundary is extended outwards along the surface method using an expansion factor:

[0086] ;

[0087] in, For point cloud normal vector field interpolation generation, This indicates the change in current in a charged body. This indicates the normal current of a charged body. To enhance the expansion factor of bi-branch optimization, Altitude correction factor;

[0088] Step S4-3: Render the core restricted area, warning area, and safe area based on the changes in the safe area in step S4-2.

[0089] Preferably, the expansion factor is trained and deployed using a two-branch optimization technique, specifically as follows:

[0090] Training phase: The dataset is a historical fault database, with noise added for augmentation. The reward function is:

[0091] ;

[0092] Deployment phase: Input real-time , Branch 1 extracts features, branch 2 outputs... ,constraint:

[0093] ;

[0094] Main strategy during online monitoring: when residual When using Backup strategy ;

[0095] Fault enforcement strategy: When At that time, forced .

[0096] Compared with the prior art, the present invention has the following improvements and advantages:

[0097] 1. By using electric field distortion compensation methods, the distortion phenomenon of point cloud data during the acquisition process is effectively reduced, improving the accuracy and reliability of point cloud data; a deformation CNN point cloud reconstruction model is constructed to accurately reconstruct the three-dimensional structure of the charged body; at the same time, by combining real-time electric field strength data with point cloud data, the safety boundary of the charged area can be more accurately delineated, reducing false detections and missed detections, and improving the timeliness and accuracy of safety warnings.

[0098] 2. The method of the present invention can adapt to the operating status of substations under different environmental conditions, and has strong robustness and adaptability. It is suitable for safety monitoring in a variety of complex scenarios. At the same time, it has broad application prospects and technical advantages in dynamic modeling of the safety area of ​​live parts in substations, and can effectively improve the safety management level and operation and maintenance efficiency of substations. Attached Figure Description

[0099] Figure 1 This is a flowchart illustrating the method of the present invention.

[0100] Figure 2 A schematic diagram of the device's security zone.

[0101] Figure 3 A schematic diagram showing the comparison results of three strategies for safe distance of charged bodies.

[0102] Figure 4 This is a schematic diagram showing the comparison results of dynamic response in the safe zone.

[0103] Figure 5 This is a schematic diagram showing the comparison results of electric field-point cloud coupling compensation accuracy. Detailed Implementation

[0104] The invention will be further summarized below with reference to the accompanying drawings.

[0105] like Figure 1 As shown, a dynamic modeling method for the safe zone of a substation's energized body based on electric field-point cloud coupling compensation is presented. The method includes the following steps:

[0106] Step S1: Collect raw point cloud data and spatial electric field gradient data of the charged body;

[0107] Step S1-1: Synchronously acquire raw point cloud data of the charged body using a rigidly connected laser scanner, a broadband electric field sensor, and an IMU attitude compensation unit. and spatial electric field gradient data The sampling frequency of the broadband electric field sensor is ≥1kHz, and the resolution of the original point cloud data of the charged body is ≤2mm.

[0108] The rigid fixed-mount laser scanner uses a carbon fiber bracket. The laser scanner and electric field sensor are precision machined to ensure that the relative displacement error is <1mm. The IMU attitude compensation unit integrates a 6-axis inertial measurement unit, model MPU-9250, with a sampling frequency of 200Hz, and outputs three-axis attitude angles (pitch angle θ, roll angle γ, yaw angle ψ) in real time to correct the deviation of the electric field gradient direction caused by equipment vibration.

[0109] The laser scanner is a Leica RTC360 model, and the wideband electric field sensor is a Narda EHP-200 model; the center-to-center distance between the laser scanner and the wideband electric field sensor is ≤10cm.

[0110] Step S1-2: Synchronization data acquisition triggered by GPS-disciplined clock source:

[0111] Step S1-2: Synchronization data acquisition triggered by GPS-disciplined clock source:

[0112] Laser scanner outputs raw point cloud data The electric field sensor collects spatial electric field gradient data as follows:

[0113] ;

[0114] in, for Gradient along the axis; for Gradient along the axis; for Gradient along the axis;

[0115] The laser scanner has a resolution of ≤2mm, a scanning rate of ≥1MHz, and a sampling frequency of ≥1kHz for the electric field sensor;

[0116] Step S1-3: Convert the Euler angles output by the IMU into a rotation matrix using a coordinate transformation matrix. ,make sure With point cloud normal vector The included angle It satisfies the formula: .

[0117] In step S2, the electric field distortion compensation method is used to compensate the original point cloud data, specifically as follows:

[0118] S2-1: Classify insulator types based on point cloud RGB features, and retrieve dielectric constant compensation coefficients from the compensation coefficient table stored in the FPGA. ;

[0119] Dielectric constant compensation coefficient Using a dielectric constant compensation coefficient table stored in the FPGA module, material type is matched in real time; the compensation calculation cycle is ≤10 ms, and the delay error is <0.1%;

[0120] when At that time, porcelain insulators were activated; When using silicone rubber bushings; porcelain insulators HSV color gamut: H180-240°, S>0.2, V 30%-60%; silicone rubber bushings: H 30-60°, S<0.5, V 50%-80%;

[0121] S2-2: Utilizing raw point cloud data Real-time electric field distortion compensation calculation is performed to generate compensated original point cloud data. :

[0122] ;

[0123] in, The raw point cloud data, The sampling interval; (Matching 1 kHz electric field sampling), FPGA pipeline computation delay ≤ 10 ms, compensation residual < 0.1%; Let be the electric field intensity gradient of the i-th element; This represents the number of insulators.

[0124] In step S3, a deformation CNN point cloud reconstruction model is constructed to reconstruct the original point cloud data and spatial electric field gradient data of the charged body. Specifically:

[0125] Step S3-1: Input spatial coordinates and electric field gradient using the input layer:

[0126] Input tensors to construct compensated original point cloud data With electric field gradient Concatenate into a 6-channel input feature tensor ;

[0127] Step S3-2: Implement basic deformation awareness from the input layer to the first execution layer:

[0128] The feature extraction module performs the DeformConv1 operation, extracting spatial features through an 11×11×11 convolution kernel. The basic deformation prediction module performs the DeformConv2 operation, predicting the basic offset using a 3×3×3 deformation convolution kernel. Output: , ;

[0129] in, The number of points in the point cloud; These are the weights of the 3D convolution kernel. This is a 3D convolution operation; It is the bias vector; It is a linear rectification activation function; This is the first stage feature map; Predict the convolution kernel for deformation; Use the Sigmoid activation function; Base offset vector;

[0130] Step S3-3: Refine the deformation field from the first execution layer to the second execution layer:

[0131] The downsampling module performs MaxPool operations. The gradient prediction module performs a 2×2×2 three-dimensional max pooling operation and a DeformConv3 operation, using a 3×3×3 deformable convolution to calculate the offset gradient. Output: , ;

[0132] in, For three-dimensional max pooling, Features after pooling; These are the kernel weights; The activation function is set to a leakage coefficient of 0.01. This is the second-stage feature map; Predict the convolution kernel for gradient; For the offset gradient field;

[0133] Step S3-4: Implement curvature-sensitive processing from the third execution layer to the fourth execution layer:

[0134] The dilated convolution module performs the DeformConv4 operation, a dilated convolution with a dilation rate of d=2, to expand the receptive field. The Jacobi calculation module is based on the displacement gradient field. Compute the Hessian operator:

[0135] ;

[0136] pass Adding a penalty term to the network loss function To prevent point cloud from being partially folded or overstretched, output , ;

[0137] in: 3D dilated convolution with dilation rate 2; This is the third-stage feature map; For Jacobian matrix fields;

[0138] Step S3-5: Implement curvature compensation from the third execution layer to the fourth execution layer:

[0139] Using the DeformConv5 convolution enhancement module The DeformConv5 operation is a 3×3×3 deformable convolution to enhance features;

[0140] The curvature tensor is calculated using the curvature compensation module, and the surface curvature is calculated using the second-order partial derivative. Output , ;

[0141] in, These are the kernel weights; This is the fourth stage feature map; For curvature tensor; The magnitude of the electric field gradient; For curvature compensation amount, This represents the geometric fit coefficient, used to compensate for normal distortion caused by the electric field gradient; It is a 1×1 convolution (dimension adjustment); This is a residual feature map;

[0142] Step S3-6: Implement field suppression from the fourth execution layer to the fifth execution layer:

[0143] Perform DeformConv6 operations using the convolutional compression module DeformConv6 uses 3×3×3 convolutions for feature compression and dimensionality reduction.

[0144] Calculate the attention to the component space using the attention module based on the electric field method. , and Threshold binding embeds safety procedure thresholds into neural network parameters, as follows: Output ;

[0145] in, It is a 3D convolution; This is the feature map for the fifth stage; This refers to the normal component of the electric field; Spatial attention weights; This is element-wise multiplication; For weighted feature maps;

[0146] Step S3-7: Deformation synthesis is achieved from the fifth execution layer to the output layer:

[0147] Perform 4x upsampling using the upsampling module, followed by trilinear interpolation. The original number of points is restored; the deformation synthesis module performs the DeformConv7 operation, which is a 3×3×3 deformation convolution, to generate the final displacement. Output ;

[0148] in, This is a 4x trilinear interpolation upsampling; For upsampling features; To output the convolution kernel, It is the hyperbolic tangent activation function; This represents the final deformation displacement;

[0149] Step S3-8: Enhance edge details along the electric field normal ;

[0150] in, This is the edge enhancement coefficient;

[0151] Step S3-9: Final Output Synthesis .

[0152] The deformation CNN point cloud reconstruction model includes a master device and slave devices, deployed on the master device of the edge computing cluster. The master device uses NVIDIA Jetson AGX Orin, 200 TOPS.

[0153] The device uses a TensorRT module to process DeformConv 1-DeformConv 4, monitoring device utilization every 5 seconds. When the equipment The migration mechanism is triggered when any of the following conditions are met:

[0154] And duration Second; $U_k(t) > 0.95;

[0155] The migration process includes the following three stages:

[0156] Freeze: Saves the layer group state, time consumed. ms;

[0157] Transmission: The compressed status data is sent via the TSN network, taking [time]. ms;

[0158] Restore: Target device loading state, time taken ms.

[0159] Edge deployment optimization is achieved using a deformation CNN point cloud reconstruction model, specifically as follows:

[0160] First, the convolutional network of the deformation CNN point cloud reconstruction model is dynamically divided into three layers based on computational cost:

[0161] Lightweight: DeformConv1-DeformConv3, computational cost TOPS;

[0162] Medium-scale: DeformConv4, computational cost TOPS;

[0163] Heavyweight: DeformConv5-DeformConv7, computational complexity TOPS;

[0164] Secondly, lightweight and medium-weight tiers are assigned to the slave device TensorR module, and heavyweight tiers are assigned to the master device to meet resource constraints. in Indicates equipment Maximum computing power;

[0165] Real-time resource monitoring, collecting the following device status:

[0166] Equipment utilization rate: ;

[0167] Load change rate: ;in, Second, This represents the utilization rate of device k at time t;

[0168] Finally, implement the dynamic degradation strategy:

[0169] When minimum resource margin TOPS:

[0170] Adjust rendering resolution: Computational scaling: scaling factor Choose according to the following rules:

[0171] .

[0172] In step S4, safe zones are defined in layers based on the reconstructed point cloud coordinates and real-time electric field intensity, specifically as follows:

[0173] Step S4-1: Based on reconstructed point cloud data Constructing implicit surfaces of charged bodies Calculate any point Minimum Euclidean distance to the implicit surface of a charged body:

[0174] :

[0175] Step S4-2: When the current change of the charged body reaches When the safety boundary is extended outwards along the surface method using an expansion factor:

[0176] ;

[0177] in, For point cloud normal vector field interpolation generation, This indicates the change in current in a charged body. This indicates the normal current of a charged body. To enhance the expansion factor of bi-branch optimization, Altitude correction factor;

[0178] The expansion factor is trained and deployed using a two-branch optimization technique, specifically:

[0179] Training phase: The dataset is a historical fault database, with noise added for augmentation. The reward function is:

[0180] ;

[0181] Deployment phase: Input real-time , Branch 1 extracts features, branch 2 outputs... ,constraint:

[0182] ;

[0183] Main strategy during online monitoring: when residual When using Backup strategy ;

[0184] Fault enforcement strategy: When At that time, forced .

[0185] Step S4-3: Render the core restricted area, warning area, and safe area based on the changes in the safe area in step S4-2.

[0186] like Figure 2 As shown, the core restricted area (>10 kV / m, distance ≥1.5 m), the warning zone (5-10 kV / m, distance ≥0.7 m), and the safe zone (<5 kV / m).

[0187] To verify the effectiveness of the method of the present invention, simulation experiments were conducted:

[0188] The experimental equipment model was a 110kV disconnector switch, with a safety reference voltage gradient of 8kV / m. The current range covered normal fluctuations to critical faults, and the electric field gradient simulated normal operation to strong electromagnetic interference. Monte Carlo simulation was used to generate 1000 sets of normal operating conditions and 200 sets of fault operating conditions.

[0189] The optimization strategies employed are compared using this scheme, a fixed strategy, and an empirical formula strategy: (e.g.) Figure 3 As shown, under normal operating conditions, the optimized redundancy distance strategy is 0.32m, the fixed strategy is 0.65m, and the empirical formula is 0.45m. Compared with the fixed strategy, the optimization rate is improved by 48.0%. Under minor faults, the optimized redundancy distance strategy is 0.78m, the fixed strategy is 0.95m, and the empirical formula is 0.82m. Compared with the fixed strategy, the optimization rate is improved by 17.9%. Under severe faults, the optimized redundancy distance strategy is 1.45m, the fixed strategy is 1.55m, and the empirical formula is 1.38m. Compared with the fixed strategy, the optimization rate is improved by 6.5%. Analysis shows that the optimized strategy has the largest redundancy reduction under normal operating conditions, but the empirical formula surpasses the fixed strategy in redundancy distance under severe faults. The optimized strategy dynamically compensates for electric field distortion, reduces redundancy distance under normal operating conditions, and has good robustness under fault scenarios.

[0190] Monte Carlo simulation was used for 5000 independent samplings, with a maximum electric field strength of 15 kV / m, a voltage gradient of 8 kV / m, and temperature cycling from 40℃ to +85℃. A 110 kV disconnector was used, and the current range covered normal fluctuations to critical faults. Figure 4As shown, based on Monte Carlo statistics, at 20ms, the safety boundary distance of this solution is 0.540±0.005m, while the pure software simulation safety boundary distance is 0.520±0.015m, representing a 3.8% improvement in safety boundary distance. At 100ms, the safety boundary distance of this solution is 0.725±0.005m, while the pure software simulation safety boundary distance is 0.600±0.015m, representing a 20.8% improvement in safety boundary distance. The noise standard deviation of this solution is 0.005, while that of the pure software simulation is 0.015, a reduction of 66.7%. This solution achieves a safety distance improvement of 0.54m in the initial stage of a fault (20ms) through dynamic boundary adjustment, and forces α=0.25 in the case of severe faults, meeting the redundancy requirements of IEC 61850 (≥1.5×0.5m). The 95% confidence interval ±0.005m corresponds to the IEEE Std C37.122.1-2020 standard, representing a 125% improvement in accuracy. Edge computing architecture implements dynamic degradation strategies to effectively reduce response latency in response to sudden loads. This solution achieves significant breakthroughs in safety distance accuracy, response speed, and fault robustness through an innovative dynamic compensation mechanism.

[0191] Experimental conditions: Original point cloud data generated from the electromagnetic model of a 500kV GIS substation; electric field sensing sampling rate of 1kHz. Comparison conditions: This scheme (electric field distortion compensation + deformation CNN reconstruction), traditional method (point cloud ICP registration compensation only), and uncompensated control group.

[0192] like Figure 5 As shown, under a strong electric field of 20kV / m, the RMSE of the point cloud after compensation for the porcelain insulator is 3.2mm, which is 32% lower than that of the traditional ICP method and 63.6% higher than that of the uncompensated scheme (8.8mm). As the electric field strength increases from 5kV / m to 20kV / m, this scheme achieves dynamic compensation of the gradient field through the curvature-sensitive module of the deformation CNN. Under the extreme condition of 20kV / m, the RMSE increase is only 1 / 3 of that of the traditional method, demonstrating excellent adaptive compensation performance of the insulator.

[0193] Injecting white noise into the electric field gradient data, the RMSE fluctuation range of the porcelain insulator was controlled within ±0.3 mm (95% confidence interval), verifying the effectiveness of the IMU attitude compensation and electric field decoupling algorithm. Note that the smaller RMSE under strong electric fields is due to the synergistic effect of enhanced regularity of electric field distortion and the material adaptive capability of the compensation algorithm, which breaks through the performance bottleneck of traditional methods under weak electric fields.

[0194] Simulation verification demonstrates the advantages of the optimization strategy in terms of redundancy distance reduction, response speed, and accuracy, providing a new technical approach for the design of substation safety protection systems.

[0195] The above description is merely an embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of the present invention should be included within the scope of the claims of the present invention.

Claims

1. A dynamic modeling method for the safe zone of a substation's energized body based on electric field-point cloud coupling compensation, characterized in that: The method includes the following steps: Step S1: Collect the raw point cloud data and spatial electric field gradient data of the charged body, specifically: Step S1-1: Synchronously acquire raw point cloud data of the charged body using a rigidly connected laser scanner, a broadband electric field sensor, and an IMU attitude compensation unit. and spatial electric field gradient data ; Step S1-2: Synchronization data acquisition triggered by GPS-disciplined clock source: Laser scanner outputs raw point cloud data The electric field sensor collects spatial electric field gradient data as follows: ； in, for Gradient along the axis; for Gradient along the axis; for Gradient along the axis; Step S1-3: Convert the Euler angles output by the IMU into a rotation matrix using a coordinate transformation matrix. ,make sure Angle with the point cloud normal vector It satisfies the formula: ; Step S2: Compensate the original point cloud data of the charged body using the electric field distortion compensation method, specifically as follows: S2-1: Classify insulator types based on point cloud RGB features, and retrieve dielectric constant compensation coefficients from the compensation coefficient table stored in the FPGA. ; Dielectric constant compensation coefficient Using a dielectric constant compensation coefficient table stored in the FPGA module, material type is matched in real time; the compensation calculation cycle is ≤10 ms, and the delay error is <0.1%; when At that time, porcelain insulators were activated; When using silicone rubber bushings; porcelain insulator HSV color gamut: H 180-240°, S>0.2, V 30%-60%; silicone rubber bushing: H 30-60°, S<0.5, V 50%-80%; S2-2: Utilizing raw point cloud data Real-time electric field distortion compensation calculation is performed to generate compensated original point cloud data. : ； in, The raw point cloud data, The sampling interval; Let be the electric field intensity gradient of the i-th element; This refers to the number of insulators. Step S3: Construct a deformation CNN point cloud reconstruction model to reconstruct the original point cloud data and spatial electric field gradient data; Step S4: Delineate safe zones based on the reconstructed point cloud coordinates and real-time electric field intensity.

2. The method for dynamic modeling of the safe zone of a substation energized body based on electric field-point cloud coupling compensation according to claim 1, characterized in that: The deformation CNN point cloud reconstruction model adopts a 7-layer network architecture, including an input layer, a first execution layer, a second execution layer, a third execution layer, a fourth execution layer, a fifth execution layer, and an output layer; The first execution layer includes a feature extraction module and a basic deformation prediction module; the second execution layer includes a downsampling module and a gradient prediction module; the third execution layer includes a dilated convolution module and a Jacobi calculation module; the fourth execution layer includes a convolution enhancement module, a curvature compensation module, a convolution compression module and an attention module; and the fifth execution layer includes an upsampling module and a deformation synthesis module. The feature extraction module uses an 11×11×11 convolution kernel to extract spatial features; the basic deformation prediction module uses a 3×3×3 deformation convolution kernel; the downsampling module uses a 2×2×2 three-dimensional max pooling; the gradient prediction module uses a 3×3×3 deformation convolution to calculate the offset gradient; the dilated convolution module uses a three-dimensional dilated convolution; and the Jacobian calculation module uses the Jacobian to calculate the Hessian operator. The convolution enhancement module uses 3×3×3 deformable convolution, the curvature compensation module calculates the curvature tensor; the convolution compression module uses 3×3×3 convolution, and the attention module is used to calculate the attention to the component space. The upsampling module performs 4x upsampling, and the deformation synthesis module uses 3×3×3 deformation convolution.

3. The method for dynamic modeling of the safe zone of a substation energized body based on electric field-point cloud coupling compensation according to claim 2, characterized in that: In step S3, a deformation CNN point cloud reconstruction model is constructed to reconstruct the original point cloud data and spatial electric field gradient data of the charged body. Specifically: Step S3-1: Input spatial coordinates and electric field gradient using the input layer: Input tensors to construct compensated original point cloud data With electric field gradient Concatenate into a 6-channel input feature tensor ; Step S3-2: Implement basic deformation awareness from the input layer to the first execution layer: The feature extraction module performs the DeformConv1 operation, extracting spatial features through an 11×11×11 convolution kernel. The basic deformation prediction module performs the DeformConv2 operation, predicting the basic offset using a 3×3×3 deformation convolution kernel. Output: , ; in, The number of points in the point cloud; These are the weights of the 3D convolution kernel. This is a 3D convolution operation; It is the bias vector; It is a linear rectification activation function; This is the first stage feature map; Predict the convolution kernel for deformation; Use the Sigmoid activation function; Base offset vector; Step S3-3: Refine the deformation field from the first execution layer to the second execution layer: The downsampling module performs MaxPool operations. The gradient prediction module performs a 2×2×2 three-dimensional max pooling operation and a DeformConv3 operation, using a 3×3×3 deformable convolution to calculate the offset gradient. Output: , ; in, For three-dimensional max pooling, Features after pooling; These are the kernel weights; The activation function is set to a leakage coefficient of 0.

01. This is the second-stage feature map; Predict the convolution kernel for gradient; For the offset gradient field; Step S3-4: Implement curvature-sensitive processing from the third execution layer to the fourth execution layer: The dilated convolution module performs the DeformConv4 operation, a dilated convolution with a dilation rate of d=2, to expand the receptive field. The Jacobi calculation module is based on the displacement gradient field. Compute the Hessian operator: ； pass Adding a penalty term to the network loss function To prevent point cloud from being partially folded or overstretched, output , ; in: 3D dilated convolution with dilation rate 2; This is the third-stage feature map; For Jacobian matrix fields; Step S3-5: Implement curvature compensation from the third execution layer to the fourth execution layer: Using the DeformConv5 convolution enhancement module The DeformConv5 operation is a 3×3×3 deformable convolution to enhance features; The curvature tensor is calculated using the curvature compensation module, and the surface curvature is calculated using the second-order partial derivative. Output , ; in, These are the kernel weights; This is the fourth stage feature map; For curvature tensor; The magnitude of the electric field gradient; For curvature compensation amount, This represents the geometric fit coefficient, used to compensate for normal distortion caused by the electric field gradient; It is a 1×1 convolution (dimension adjustment); This is a residual feature map; Step S3-6: Implement field suppression from the fourth execution layer to the fifth execution layer: Perform DeformConv6 operations using the convolutional compression module DeformConv6 uses 3×3×3 convolutions for feature compression and dimensionality reduction. Calculate the attention to the component space using the attention module based on the electric field method. , and Threshold binding embeds safety procedure thresholds into neural network parameters, as follows: Output ; in, It is a 3D convolution; This is the feature map for the fifth stage; This refers to the normal component of the electric field; Spatial attention weights; This is element-wise multiplication; For weighted feature maps; Step S3-7: Deformation synthesis is achieved from the fifth execution layer to the output layer: Perform 4x upsampling using the upsampling module, followed by trilinear interpolation. The original number of points is restored; the deformation synthesis module performs the DeformConv7 operation, which is a 3×3×3 deformation convolution, to generate the final displacement. Output ; in, This is a 4x trilinear interpolation upsampling; For upsampling features; To output the convolution kernel, It is the hyperbolic tangent activation function; This represents the final deformation displacement; Step S3-8: Enhance edge details along the electric field normal ; in, This is the edge enhancement coefficient; Step S3-9: Final Output Synthesis .

4. The method for dynamic modeling of the safe zone of a substation energized body based on electric field-point cloud coupling compensation according to claim 3, characterized in that: The deformation CNN point cloud reconstruction model includes a master device and a slave device, with the slave device employing a TensorR module. Edge deployment optimization is achieved using a deformation CNN point cloud reconstruction model, specifically as follows: First, the convolutional network of the deformation CNN point cloud reconstruction model is dynamically divided into three layers based on computational cost: Lightweight: DeformConv1-DeformConv3, computational cost TOPS; Medium-scale: DeformConv4, computational cost TOPS; Heavyweight: DeformConv5-DeformConv7, computational complexity TOPS; Secondly, lightweight and medium-weight tiers are assigned to the slave device TensorR module, and heavyweight tiers are assigned to the master device to meet resource constraints. in Indicates equipment Maximum computing power; Real-time resource monitoring, collecting the following device status: Equipment utilization rate: ; Load change rate: ;in, Second, This represents the utilization rate of device k at time t; Finally, implement the dynamic degradation strategy: When minimum resource margin TOPS: Adjust rendering resolution: Computational scaling: ; scaling factor Choose according to the following rules: 。 5. The method for dynamic modeling of the safe zone of a substation energized body based on electric field-point cloud coupling compensation according to claim 4, characterized in that: The deformation CNN point cloud reconstruction model has migration triggering conditions in achieving edge deployment optimization, specifically: When the device The migration mechanism is triggered when any of the following conditions are met: And duration Second; $U_k(t) > 0.95; The migration process includes the following three stages: Freeze: Saves the layer group state, time consumed. ms; Transmission: The compressed status data is sent via the TSN network, taking [time]. ms; Restore: Target device loading state, time taken ms.

6. The method for dynamic modeling of the safe zone of a substation energized body based on electric field-point cloud coupling compensation according to claim 1, characterized in that: In step S4, the safe zone is delineated in layers based on the reconstructed point cloud coordinates and real-time electric field intensity, specifically as follows: Step S4-1: Based on reconstructed point cloud data Constructing implicit surfaces of charged bodies Calculate any point Minimum Euclidean distance to the implicit surface of a charged body: ： Step S4-2: When the current change of the charged body reaches When the safety boundary is extended outwards along the surface method using an expansion factor: ； in, For point cloud normal vector field interpolation generation, This indicates the change in current in a charged body. This indicates the normal current of a charged body. To enhance the expansion factor of bi-branch optimization, Altitude correction factor; Step S4-3: Render the core restricted area, warning area, and safe area based on the changes in the safe area in step S4-2.

7. The method for dynamic modeling of the safe zone of a substation energized body based on electric field-point cloud coupling compensation according to claim 6, characterized in that: The expansion factor is trained and deployed using a dual-branch optimization technique, specifically: Training phase: The dataset is a historical fault database, with noise added for augmentation. The reward function is: ； Deployment phase: Input real-time , Branch 1 extracts features, branch 2 outputs... ,constraint: ； Main strategy during online monitoring: when residual When using ; Backup strategy ; Fault enforcement strategy: When At that time, forced .

Images