Automatic efficient modeling method and system for electricity utilization inspection scene

Abstract

The invention relates to the technical field of electric power system digitization, in particular to an automatic efficient modeling method and system for an electricity utilization inspection scene. The automatic efficient modeling method for the electricity utilization inspection scene comprises the four steps of multi-source data collection, dynamic semantic modeling, intelligent modeling decision making and augmented reality rendering, the electricity utilization inspection scene modeling system comprises a data collection module, an intelligent modeling engine, an augmented reality terminal and edge computing equipment, and data interaction is achieved among the modules through a gRPC protocol. According to the method, an L4-level multi-physical field model is generated through unmanned aerial vehicle cluster collaborative collection and multi-source data millisecond-level fusion in combination with a dynamic semantic network and reinforcement learning resource allocation; edge calculation and dynamic LOD rendering are adopted to realize equipment internal perspective and multi-user collaborative labeling, quality closed-loop verification is matched, the bottlenecks of low modeling efficiency, multi-physical field data missing, high interaction delay and the like in the prior art are overcome, and the overall efficiency is improved by 80%.

Description

Technical Field

[0001] The present invention relates to the technical field of digitalization of power systems, and particularly relates to an automated and efficient modeling method and system for electricity consumption inspection scenarios. Background Art

[0002] Currently, the power industry has an increasingly urgent need for digital modeling of electricity consumption inspection scenarios. Especially in scenarios such as distribution room inspections, electricity theft detection, and equipment status monitoring, traditional technologies face significant bottlenecks. Existing modeling methods mainly rely on manual measurement and static three-dimensional reconstruction technologies, and their limitations are reflected in the following aspects:

[0003] First, the efficiency of data collection and modeling is low. Traditional manual measurement relies on equipment such as total stations to collect data point by point, and the time-consuming for a single modeling is more than 8 hours, which is difficult to meet the real-time requirements. Although existing automated modeling technologies use a single sensor to collect data, they do not solve the problem of the fusion of multi-source heterogeneous data. The operating parameters of power equipment and three-dimensional spatial models are often stored independently, resulting in the lack of dynamic association between the modeling results and the equipment status and spatial position. In addition, the convergence time of the flight path planning algorithm of existing UAV aerial photography modeling technologies generally exceeds 60 seconds, and the point cloud stitching error is often higher than 2 cm, which is difficult to meet the modeling requirements of millimeter-level accuracy for power equipment.

[0004] Second, the ability of model dynamic update and scene understanding is insufficient. Traditional three-dimensional models are mostly static geometric structures and cannot reflect dynamic characteristics such as equipment deterioration and environmental changes. For example, key state changes such as the rusting of switch cabinets and the temperature rise of cable joints cannot be mapped to the model in real time. Although existing technologies can construct the logical topology of equipment, they lack the semantic modeling ability for the multi-dimensional relationship of equipment-environment-operation conditions.

[0005] Third, the limitations of visualization and interaction technologies. Traditional AR rendering systems mostly use general game engines, and the rendering frame rate is usually lower than 30 fps, and the coupled visualization of power physical fields cannot be achieved. In the scenario of multi-person collaborative operation, due to the communication protocol delay and uneven rendering load in existing systems, problems such as operation jamming and annotation misalignment often occur. Summary of the Invention

[0006] The purpose of the present invention is to overcome the above problems and provide an automated and efficient modeling method for electricity consumption inspection scenarios. To achieve the above purpose, the present invention adopts the following technical solutions:

[0007] An automated and efficient modeling method for electricity consumption inspection scenarios includes the following steps:

[0008] Step S1: Multi-source data collection: Synchronously obtain the three-dimensional point cloud data of the scene through the oblique photography unit and the laser scanning device carried by the UAV cluster, and the point cloud density is not less than 500 points / m 2, and access the real-time operation parameters of the power Internet of Things sensing terminal through the IEC 61850 protocol;

[0009] Step S2: Dynamic semantic modeling: Construct a power knowledge graph containing device types, operating states, and environmental parameters, extract entity relationships from device manuals based on the BERT model, establish a mapping relationship between the three-dimensional space coordinate system and the device logical topology, and generate a dynamic semantic network with time series characteristics;

[0010] Step S3: Intelligent modeling decision: Dynamically allocate modeling resources based on the reinforcement learning algorithm, initialize network parameters using the transfer learning pre-training strategy, generate scenario models with L1-L4 accuracy levels, and the geometric error of the L4-level model is ≤0.1 mm and includes electromagnetic-temperature multi-physical field coupling data;

[0011] Step S4: Augmented reality rendering: Achieve perspective visualization of the internal structure of the device through a heterogeneous rendering engine, use the dynamic LOD control algorithm to maintain a rendering frame rate ≥90 fps, and support multi-user collaborative annotation operations.

[0012] Furthermore, in step S1, the drone cluster adopts an improved TSP trajectory planning algorithm, and the conditions are as follows: the cluster scale is 3-7 drones, the flight altitude is 10-15 m, the overlap degree of adjacent flight routes is 30%-40%, the convergence time is shortened to ≤30 s by introducing a taboo search mechanism, and the point cloud stitching adopts a feature point matching algorithm with a matching error ≤2 cm.

[0013] Furthermore, in step S2, the construction of the dynamic semantic network includes a device state transition matrix and an environmental association model; the device state transition matrix defines an 8-dimensional device deterioration feature vector and a state transition probability; the environmental association model establishes an association function between temperature, humidity, vibration parameters and the device state as:

[0014] f(x) = α·e β(T-T0) +γ·log(1 + δV);

[0015] where α, β, γ, δ are material characteristic coefficients, T is temperature, and V is vibration intensity.

[0016] Furthermore, in step S3, the reinforcement learning algorithm adopts a deep deterministic policy gradient framework, its state space dimension is 15-dimensional, the action space dimension is 8-dimensional, and the calculation formula of the reward function is as follows:

[0017]

[0018] where Q ∈ [0,1] is the modeling quality score, T is the actual time consumption, and C is the resource consumption coefficient.

[0019] Further, in step S4, the dynamic LOD control algorithm satisfies:

[0020]

[0021] where D is the observation distance, and the number of model patches corresponding to each LOD level is: L1 ≤ 5k, L4 ≥ 200k.

[0022] Further, an electricity inspection scenario modeling system includes: a data acquisition module, an intelligent modeling engine, an augmented reality terminal, and an edge computing device. Data interaction is realized between the modules through the gRPC protocol; the data acquisition module includes a drone cluster control unit, a laser scanning synchronization device, and an Internet of Things perception data repeater; the intelligent modeling engine is built-in with a multi-source data fusion algorithm, a dynamic semantic network generator, and a modeling quality evaluation unit; the augmented reality terminal is equipped with a perspective rendering shader, a collaborative operation communication interface, and a gesture recognition sensor; the edge computing device integrates an NVIDIA A6000 GPU and a 5G communication module, and the computing latency ≤ 20ms.

[0023] Further, the data fusion algorithm adopts an improved iterative closest point registration technique, including point cloud preprocessing, a rough registration stage, a fine registration stage, and outlier removal;

[0024] The point cloud preprocessing reduces the original point cloud resolution to 1 - 3 cm through voxel grid filtering 3 , and retains a feature point density ≥ 200 points / cm 2 ;

[0025] In the rough registration stage, the FPFH feature descriptor is used for initial matching, and the rotation matrix R and the translation vector t are calculated. The formula is as follows;

[0026]

[0027] where the number of iterations ≤ 5 times;

[0028] In the fine registration stage, the LM optimization algorithm is applied to adjust the transformation matrix to satisfy the final registration error ≤ 0.5 cm;

[0029] The outlier removal removes the mismatched point pairs based on Mahalanobis distance detection (threshold σ = 2.5).

[0030] Further, the modeling quality evaluation unit includes a geometric accuracy detector, a physical field verification module, and a real-time scoring system;

[0031] The geometric accuracy detector uses the SIFT key point matching algorithm to compare the model with the actual point cloud, and defines the geometric error index as:

[0032]

[0033] Where M is the model coordinate, S is the scanning coordinate, and Davg is the minimum feature size of the device;

[0034] The physical field verification module measures the field strength distribution by deploying electromagnetic field probes, and calculates the correlation coefficient between the simulated value and the measured value. The formula is as follows:

[0035]

[0036] The real-time scoring system constructs a quality evaluation model Q = 0.6(1 - Eg) + 0.4ρ, generates a quality report every 30 seconds, and its output quality index Q ∈ [0, 1].

[0037] The advantages of the present invention are as follows:

[0038] 1. The present invention realizes the efficient fusion of multi-source data and dynamic precise modeling. Through the collaborative scanning of UAV clusters and laser point cloud technology, combined with an improved trajectory planning algorithm, the modeling efficiency of complex scenes is increased to 8 times that of traditional methods, and the modeling time is shortened to 40 minutes / 100m 2 . An innovative dynamic semantic network is constructed, integrating device operation parameters, environmental data and 3D models, supporting dynamic updates of distribution rooms, electricity theft scenarios, etc. The modeling accuracy of key devices reaches 0.1mm, and the abnormal recognition accuracy rate exceeds 98%.

[0039] 2. The present invention breaks through the bottlenecks of intelligent modeling decision-making and resource optimization. It uses a reinforcement learning algorithm to dynamically allocate computing resources, adjusts parameters in real time based on device status and modeling requirements, and realizes a second-level response on edge computing devices. Through a hierarchical modeling strategy (L1-L4) and a quality evaluation system, while ensuring accuracy, the GPU video memory occupancy is reduced by 37%, and the temperature field simulation error is controlled within 1.5°C, significantly improving the modeling efficiency of complex scenes.

[0040] 3. The present invention constructs a power-specific augmented reality interaction system, realizes the internal perspective of devices and electromagnetic field visualization based on a heterogeneous rendering engine, dynamically adjusts the model detail level, and the rendering frame rate is stable above 90fps. It supports multi-user collaborative annotation and gesture operations, with a delay of less than 30ms and a positioning error of less than 1mm. It can quickly reconstruct electricity theft scenarios and real-time mark abnormal points, with an efficiency 5 times higher than that of traditional manual inspections, providing an immersive solution for power inspections. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] The drawings forming a part of this application are used to provide a further understanding of this application, making other features, objects, and advantages of this application more obvious. The schematic embodiments and descriptions of the drawings of this application are used to explain this application and do not constitute an improper limitation of this application.

[0042] In the accompanying drawings:

[0043] Figure 1 is a flowchart of the automated and efficient modeling method for the electricity inspection scenario in Embodiment 1.

[0044] Figure 2 is a module interaction diagram of the automated and efficient modeling system for the electricity inspection scenario in Embodiment 1. Detailed implementation manners

[0045] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. Usually, the components of the embodiments of the present invention described and illustrated in the accompanying drawings here can be arranged and designed in various different configurations.

[0046] The present invention will be described in detail and specifically below through specific embodiments to better understand the present invention. However, the following embodiments do not limit the protection scope of the present invention.

[0047] Embodiment 1

[0048] As Figure 1-2 shown, the automated and efficient modeling method for the electricity inspection scenario includes the following steps:

[0049] Step S1: Multi-source data collection: Synchronously obtain the three-dimensional point cloud data of the scene through the oblique photography unit and the laser scanning device carried by the UAV cluster, and the point cloud density is not less than 500 points / m 2 , and access the real-time operation parameters of the power IoT sensing terminal through the IEC 61850 protocol;

[0050] Through the improved TSP trajectory planning algorithm and the cooperative control of the UAV cluster (scale of 3 - 7), the trajectory planning time is shortened to ≤30 s, and a flight altitude of 10 - 15 m achieves a flight line overlap of 30% - 40%, ensuring that the point cloud density ≥500 points / m 2 . Combining the IoT sensing data accessed by the laser scanning device and the IEC 61850 protocol, synchronously obtain the device operation parameters, solve the problem of traditional inspection data fragmentation, and the point cloud stitching error ≤2 cm, providing a complete data basis for high-precision modeling.

[0051] Step S2: Dynamic semantic modeling: Construct a power knowledge graph including device types, operating states, and environmental parameters, extract entity relationships from device manuals based on the BERT model, establish the mapping relationship between the three-dimensional space coordinate system and the device logical topology, and generate a dynamic semantic network with time series characteristics;

[0052] Extract entity relationships from device manuals based on the BERT model, construct a power knowledge graph, and map the device logical topology through a three-dimensional coordinate system to form a dynamic semantic network containing time series features. Define an 8-dimensional device deterioration feature vector and a state transition probability matrix, and combine the environmental correlation function to quantitatively analyze the impact of temperature, humidity, and vibration on the device state, providing a computable model for device life prediction.

[0053] Step S3: Intelligent modeling decision: Dynamically allocate modeling resources based on the reinforcement learning algorithm, initialize network parameters using the transfer learning pre-training strategy, generate scenario models with accuracy levels L1 - L4, where the geometric error of the L4-level model is ≤0.1mm and it includes electromagnetic-temperature multi-physics field coupling data;

[0054] Adopt the deep deterministic policy gradient framework to dynamically allocate computing resources with a 15-dimensional state space and an 8-dimensional action space. The reward function drives the modeling time to be reduced by 40%. Generate L1 - L4 level models through transfer learning pre-trained parameters, where the geometric error of the L4-level model is ≤0.1mm, and integrate electromagnetic field and temperature field coupling data to support the localization of multi-physics field anomalies inside the device.

[0055] Step S4: Augmented reality rendering: Achieve the perspective visualization of the internal structure of the device through a heterogeneous rendering engine, adopt the dynamic LOD control algorithm to maintain the rendering frame rate ≥90fps, and support multi-user collaborative annotation operations.

[0056] The heterogeneous rendering engine uses the dynamic LOD control algorithm to hierarchically load model details according to the viewing distance (D) and the device status: When D ≤ 2m and the device is abnormal, the L4-level model is enabled, and in other scenarios, it switches to the L1-level, maintaining the rendering frame rate ≥90fps on the mobile device. Combining the gesture recognition sensor and the collaborative annotation interface, it realizes multi-user operation without delay, supporting the perspective of the internal structure of the device and fault marking.

[0057] Furthermore, in step S1, the drone swarm adopts an improved TSP trajectory planning algorithm, with the following conditions: the swarm size is 3 - 7 drones, the flight altitude is 10 - 15m, the overlap degree of adjacent flight paths is 30% - 40%, the convergence time is shortened to ≤30s by introducing a tabu search mechanism, and feature point matching algorithm is used for point cloud stitching with a matching error ≤2cm.

[0058] Furthermore, in step S2, the construction of the dynamic semantic network includes a device state transition matrix and an environmental correlation model; the device state transition matrix defines an 8-dimensional device deterioration feature vector and a state transition probability; the environmental correlation model establishes the correlation function between temperature, humidity, vibration parameters and the device state as:

[0059] f(x) = α·e β(T-T0) +γ·log(1 + δV);

[0060] Among them, α, β, γ, and δ are material characteristic coefficients, T is temperature, and V is vibration intensity.

[0061] Furthermore, in step S3, the reinforcement learning algorithm adopts a deep deterministic policy gradient framework, with a state space dimension of 15 and an action space dimension of 8. The calculation formula of the reward function is as follows:

[0062]

[0063] Among them, Q ∈ [0, 1] is the modeling quality score, T is the actual time consumption, and C is the resource consumption coefficient.

[0064] Furthermore, in step S4, the dynamic LOD control algorithm satisfies:

[0065]

[0066] Among them, D is the observation distance, and the number of model patches corresponding to each LOD level is: L1 ≤ 5k, L4 ≥ 200k.

[0067] Furthermore, an electricity inspection scenario modeling system includes: a data acquisition module, an intelligent modeling engine, an augmented reality terminal, and an edge computing device. Data interaction between modules is achieved through the gRPC protocol; the data acquisition module includes a drone cluster control unit, a laser scanning synchronization device, and an Internet of Things perception data repeater; the intelligent modeling engine is built-in with a multi-source data fusion algorithm, a dynamic semantic network generator, and a modeling quality evaluation unit; the augmented reality terminal is equipped with a perspective rendering shader, a collaborative operation communication interface, and a gesture recognition sensor; the edge computing device integrates an NVIDIA A6000 GPU and a 5G communication module, with a computing latency ≤ 20ms.

[0068] The data acquisition module consists of a drone cluster control unit, a laser scanning synchronization device, and an Internet of Things perception data repeater. Its core function is to achieve efficient synchronous acquisition of multi-source data. The drone cluster adopts an improved TSP trajectory planning algorithm to complete the collaborative trajectory planning of 3 - 7 drones within 30 seconds, synchronously trigger the oblique photography and laser scanning devices, and generate a three-dimensional point cloud with a density ≥ 500 points / m 2 and perform stitching through a feature point matching algorithm. At the same time, the Internet of Things perception data repeater accesses parameters such as current, voltage, and temperature of power equipment in real time based on the IEC 61850 protocol to ensure millisecond-level synchronization of spatial data and device status. The data acquisition efficiency is 80% higher than that of traditional manual inspections, and the time alignment error between the point cloud and operating parameters ≤ 0.1 second, providing high-precision input for modeling.

[0069] The terminal is equipped with a perspective rendering shader, a collaborative operation interface, and a gesture recognition sensor, focusing on high-frame-rate visualization and interaction optimization. The dynamic LOD control algorithm hierarchically loads model details according to the viewing distance (D) and device status: when D ≤ 2m and the device is abnormal, the L4-level model perspective display is enabled to show the internal structure; in other scenarios, it switches to L1 level, and combines with a heterogeneous rendering engine to maintain a frame rate ≥ 90fps. The collaborative operation interface supports multi-user real-time annotation, and the annotation data is synchronized to the edge device through the 5G module; the gesture recognition sensor enables contactless operation through infrared optical capture. The rendering performance is improved by 70%, the interaction response speed is improved by 90%, and multi-person collaborative fault marking is supported.

[0070] The edge computing device integrates an NVIDIA A6000 GPU and a 5G communication module, undertaking real-time computing and data transfer tasks. The GPU parallelly processes point cloud registration and semantic modeling, and the 5G module ensures real-time data synchronization among the drone, the modeling engine, and the AR terminal. The built-in reinforcement learning model dynamically allocates CPU / GPU resources to optimize the modeling time consumption and resource consumption. The end-to-end modeling time is compressed from 30 minutes to ≤ 5 minutes, meeting the on-site real-time decision-making requirements.

[0071] Furthermore, the data fusion algorithm adopts an improved iterative closest point registration technique, including point cloud preprocessing, rough registration stage, fine registration stage, and outlier removal;

[0072] The point cloud preprocessing reduces the original point cloud resolution to 1 - 3cm through voxel grid filtering 3 , retaining a feature point density ≥ 200 points / cm 2 ;

[0073] In the rough registration stage, the FPFH feature descriptor is used for initial matching, calculating the rotation matrix R and the translation vector t, with the formula as follows;

[0074]

[0075] where the number of iterations ≤ 5 times;

[0076] In the fine registration stage, the LM optimization algorithm is applied to adjust the transformation matrix to meet the final registration error ≤ 0.5cm;

[0077] The outlier removal removes mismatched point pairs based on Mahalanobis distance detection (threshold σ = 2.5).

[0078] Furthermore, the modeling quality assessment unit includes a geometric accuracy detector, a physical field verification module, and a real-time scoring system;

[0079] The geometric accuracy detector uses the SIFT key point matching algorithm to compare the model with the actual point cloud, and defines the geometric error index as:

[0080]

[0081] Where M is the model coordinate, S is the scanning coordinate, and Davg is the minimum feature size of the device;

[0082] The physical field verification module measures the field strength distribution by deploying electromagnetic field probes and calculates the correlation coefficient between the simulated value and the measured value. The formula is as follows:

[0083]

[0084] The real-time scoring system constructs a quality evaluation model Q=0.6(1-Eg)+0.4ρ, generates a quality report every 30 seconds, and outputs a quality index Q∈[0,1].

[0085] The intelligent modeling engine integrates multi-source data fusion algorithm, dynamic semantic network generator and modeling quality assessment unit to realize the automatic generation from raw data to multi-precision model. The data fusion algorithm completes point cloud registration through improved ICP technology; the dynamic semantic network generator parses the equipment manual based on the BERT model, builds a spatiotemporal semantic model including equipment degradation feature vector (8 dimensions) and environmental correlation function, and quantitatively analyzes the impact of the environment on the equipment status; the quality assessment unit verifies through SIFT key point matching and electromagnetic field measurement, outputs the comprehensive quality index Q every 30 seconds, and drives the dynamic optimization of the model. Supports L4 model generation, and the quality assessment false alarm rate is ≤0.5%.

[0086] The specific embodiments of the present invention are described in detail above, but they are only examples, and the present invention is not equivalent to the specific embodiments described above. For those skilled in the art, any equivalent modifications and substitutions made to the present invention are also within the scope of the present invention. Therefore, the equalization changes and modifications made without departing from the spirit and scope of the present invention should be included in the scope of the present invention.

Claims

1. An automated and efficient modeling method for electricity inspection scenarios, characterized in that: It includes the following steps: Step S1: Multi-source data acquisition: Synchronously obtain the 3D point cloud data of the scene through the oblique photography unit and the laser scanning device carried by the UAV cluster, and the point cloud density is not less than 500 points / m 2 , and access the real-time operation parameters of the power IoT sensing terminal through the IEC 61850 protocol; Step S2: Dynamic semantic modeling: Construct a power knowledge graph including device type, operating status, and environmental parameters. Extract entity relationships from device manuals based on the BERT model, establish the mapping relationship between the three-dimensional space coordinate system and the device logical topology, and generate a dynamic semantic network with time series characteristics; Step S3: Intelligent modeling decision-making: Dynamically allocate modeling resources based on the reinforcement learning algorithm, initialize network parameters using the transfer learning pre-training strategy, generate scenario models with L1-L4 accuracy levels, where the geometric error of the L4-level model is ≤0.1 mm and it includes electromagnetic-temperature multi-physical field coupling data; Step S4: Augmented reality rendering: Achieve the perspective visualization of the internal structure of the device through a heterogeneous rendering engine, use the dynamic LOD control algorithm to maintain the rendering frame rate ≥90 fps, and support multi-user collaborative annotation operations.

2. The automated and efficient modeling method for power consumption inspection scenarios according to claim 1, wherein: In step S1, the drone cluster adopts an improved TSP trajectory planning algorithm, and the conditions are as follows: the cluster scale is 3-7 drones, the flight altitude is 10-15 m, the overlap degree of adjacent flight lines is 30%-40%, and the convergence time is shortened to ≤30 s by introducing a taboo search mechanism. Feature point matching algorithm is used for point cloud stitching, and the matching error is ≤2 cm.

3. The automated and efficient modeling method for power consumption inspection scenarios according to claim 2, wherein In step S2, the construction of the dynamic semantic network includes a device state transition matrix and an environmental association model; the device state transition matrix defines an 8-dimensional device deterioration feature vector and state transition probability; the environmental association model establishes the association function between temperature, humidity, vibration parameters and device state as: where α, β, γ, δ are material characteristic coefficients, T is temperature, and V is vibration intensity.

4. The automated and efficient modeling method for power consumption inspection scenarios according to claim 3, characterized in that: In step S3, the reinforcement learning algorithm adopts a deep deterministic policy gradient framework, its state space dimension is 15, the action space dimension is 8, and the calculation formula of the reward function is as follows: where Q ∈ [0,1] is the modeling quality score, T is the actual time consumption, and C is the resource consumption coefficient.

5. The automated and efficient modeling method for power consumption inspection scenarios according to claim 4, characterized in that: In step S4, the dynamic LOD control algorithm satisfies: where D is the observation distance, and the number of model patches corresponding to each LOD level is: L1 ≤ 5k, L4 ≥ 200k.

6. An electricity inspection scenario modeling system for implementing the method according to any one of claims 1-5, characterized in that, It includes: A data acquisition module, an intelligent modeling engine, an augmented reality terminal, and an edge computing device. Data interaction between modules is realized through the gRPC protocol; the data acquisition module includes a drone cluster control unit, a laser scanning synchronization device, and an IoT perception data repeater; the intelligent modeling engine is built-in with a multi-source data fusion algorithm, a dynamic semantic network generator, and a modeling quality evaluation unit; the augmented reality terminal is equipped with a perspective rendering shader, a collaborative operation communication interface, and a gesture recognition sensor; the edge computing device integrates an NVIDIA A6000 GPU and a 5G communication module, and the computing delay is ≤20 ms.

7. The power consumption inspection scenario modeling system according to claim 6, characterized in that The data fusion algorithm adopts an improved iterative closest point registration technique, including point cloud preprocessing, rough registration stage, fine registration stage, and outlier removal; The above point cloud preprocessing reduces the resolution of the original point cloud to 1-3 cm through voxel grid filtering 3 , and retains the feature point density ≥ 200 points / cm 2 ; In the rough registration stage, the FPFH feature descriptor is used for initial matching, and the rotation matrix R and translation vector t are calculated, and the formula is as follows; where the number of iterations is ≤5 times; In the fine registration stage, the LM optimization algorithm is applied to adjust the transformation matrix to meet the final registration error ≤ 0.5 cm; The outlier rejection is based on Mahalanobis distance detection (threshold σ = 2.5) to remove the mismatched point pairs.

8. The power consumption inspection scenario modeling system according to claim 6, characterized in that The modeling quality evaluation unit includes a geometric accuracy detector, a physical field verification module, and a real-time scoring system; The geometric accuracy detector uses the SIFT key point matching algorithm to compare the model with the actual point cloud, and defines the geometric error index as: where M is the model coordinate, S is the scanning coordinate, and Davg is the minimum feature size of the device; The physical field verification module measures the field strength distribution by deploying electromagnetic field probes and calculates the correlation coefficient between the simulated value and the measured value. The formula is as follows: The real-time scoring system constructs a quality evaluation model Q = 0.6(1 - Eg) + 0.4ρ, generates a quality report every 30 seconds, and its output quality index Q ∈ [0, 1].

Images