Substation camera automatic layout optimization system based on three-dimensional real scene modeling

Abstract

The application relates to the technical field of substation automation inspection, in particular to a substation camera automatic distribution optimization system based on three-dimensional real scene modeling; the system comprises data acquisition and modeling, task constraint setting, space-time light shadow simulation, imaging evaluation, multi-target cooperative optimization and blind area analysis and feedback modules; the system generates virtual light shadow distribution by constructing a three-dimensional semantic real scene model and combining a light ray tracking algorithm; the core is to construct a target function according to virtual imaging evaluation feature recognizable and combining distribution cost, and to iteratively optimize camera coordinates and postures through heuristic search; when a blind area is found, the system will feed back the blind area features to the optimization module, locally increase nodes and re-optimize; the application realizes optimal balance of recognition accuracy and hardware cost, can accurately predict dynamic light and shadow and form a blind area compensation closed loop, and guarantees all-weather and dead angle-free monitoring of the substation.

Description

Automated Substation Camera Deployment Optimization System Based on 3D Real-Scene Modeling Technical Field

[0001] This invention relates to the field of substation automated inspection technology, specifically a substation camera automated deployment optimization system based on 3D real-scene modeling. Background Technology

[0002] In the current 24 / 7 monitoring and automated inspection environment of substations, reasonable camera deployment is fundamental to ensuring the effective acquisition and analysis of the status of target equipment. However, existing camera deployment schemes have significant shortcomings when dealing with the complex environment of substations. Due to the complex geometric topology and diverse material reflection properties of target equipment in substations, traditional static mapping methods cannot accurately reflect the reflection and refraction behavior of light on different material surfaces, easily leading to the loss of high-light characteristics. This results in the inability to effectively predict dynamic spatiotemporal light and shadow changes in actual planning, easily leading to invalid deployments due to strong light blindness and optical blind spots caused by complex obstructions. In the absence of quantitative evaluation of image features, existing solutions often rely on stacking camera nodes to try to improve monitoring coverage. This not only fails to reliably meet the requirements of high-precision recognition, but also multiplies hardware and construction costs, resulting in serious waste of resources.

[0003] Therefore, how to accurately avoid ineffective blind spots under the complex dynamic spatiotemporal light and shadow constraints of substations, and achieve the optimal balance between ensuring high recognition rate of inspection images and global camera deployment cost, has become an urgent technical problem to be solved. Summary of the Invention

[0004] The purpose of this invention is to provide an automated substation camera deployment optimization system based on 3D real-scene modeling. Addressing the problems of insufficient dynamic light and shadow prediction and cost imbalance in existing deployment schemes under complex environments, this invention aims to avoid ineffective blind spots by accurately predicting light and shadow, and to achieve an optimal balance in global deployment costs while ensuring high recognition rates for inspection images. To achieve the above objectives, this invention adopts the following technical solution:

[0005] An automated deployment optimization system for substation cameras based on 3D real-scene modeling includes a cloud platform, which is equipped with a data acquisition and modeling module, a task constraint setting module, a spatiotemporal light and shadow simulation module, an imaging evaluation module, and a multi-objective collaborative optimization module.

[0006] The data acquisition and modeling module is used to acquire three-dimensional point cloud data of the target substation area, perform semantic segmentation on the three-dimensional point cloud data, extract the spatial geometric features and material reflection properties of the target equipment based on the semantic segmentation results, and construct a three-dimensional semantic reality model based on the spatial geometric features and material reflection properties.

[0007] The task constraint setting module is used to obtain the inspection task type of the target device, and allocate the corresponding pixel density threshold, observation angle range and occlusion ratio upper limit according to the inspection task type to generate a task semantic constraint set.

[0008] The spatiotemporal light and shadow simulation module is used to acquire the latitude and longitude data and time period of the target substation area. Based on the latitude and longitude data and time period, a solar trajectory model is constructed. Combined with the three-dimensional semantic real scene model and the task semantic constraint set, virtual light and shadow distribution data for different time periods are generated through ray tracing algorithm.

[0009] The imaging evaluation module is used to set the initial monitoring node in the three-dimensional semantic reality model, generate virtual imaging data corresponding to the viewpoint of the initial monitoring node by combining virtual light and shadow distribution data, input the virtual imaging data into the image feature evaluation algorithm, and output the feature recognizability score of the initial monitoring node.

[0010] The multi-objective collaborative optimization module is used to obtain the feature identifiability score and deployment cost parameters of the initial monitoring nodes. Based on the feature identifiability score and deployment cost parameters, an objective function is constructed. The spatial coordinates and attitude parameters of the initial monitoring nodes are iteratively optimized through a heuristic search algorithm to output the target deployment scheme.

[0011] Furthermore, the process of constructing a 3D semantic reality model includes:

[0012] Acquire 3D point cloud data of various target devices within the target substation area, and use clustering algorithms to separate the 3D point cloud data into independent device point cloud sets;

[0013] Extract the geometric topology of the independent device point cloud, match the geometric topology with a preset device component library containing the mapping relationship between device geometric topology and component semantic tags, and semantically annotate the independent device point cloud based on the matching results to obtain component semantic tags;

[0014] Obtain the surface material parameters corresponding to the semantic tags of the components, map the surface material parameters and component semantic tags to the geometric topology, and generate a 3D semantic reality model.

[0015] Furthermore, the process of generating virtual light and shadow distribution data at different time periods using ray tracing algorithms includes:

[0016] Extract several discrete time nodes within the time period, and obtain the solar incident ray vector corresponding to each discrete time node based on the solar trajectory model;

[0017] The solar incident ray vector is projected onto the 3D semantic reality model, and the intersection point between the solar incident ray vector and the surface of the target device in the 3D semantic reality model is extracted.

[0018] The reflected ray vector and refracted ray vector are calculated based on the material reflection properties at the intersection point. The light intensity value of the target device surface is obtained based on the reflected ray vector, refracted ray vector and solar incident ray vector. The light intensity values ​​of each discrete time node are collected to generate virtual light and shadow distribution data.

[0019] Furthermore, the process of generating virtual imaging data corresponding to the viewpoint of the initial monitoring node includes:

[0020] In the three-dimensional continuous space of the three-dimensional semantic reality model, the spatial coordinates and attitude parameters of several initial monitoring nodes are randomly generated;

[0021] Starting from spatial coordinates, a view frustum bounding box is constructed by combining attitude parameters, and the surface data of the 3D semantic reality model within the view frustum bounding box is extracted;

[0022] The surface data of the 3D semantic reality model within the view frustum bounding box is fused and rendered with the virtual light and shadow distribution data to generate a virtual 2D pixel matrix corresponding to the viewpoint, and the virtual 2D pixel matrix is ​​marked as virtual imaging data.

[0023] Furthermore, the process of outputting the feature identifiability score of the initial monitoring node includes:

[0024] The virtual imaging data is input into the image feature evaluation algorithm, and the image feature evaluation algorithm outputs the pixel density value of the target device, the proportion of highlight overflow area, and the proportion of shadow occlusion area in the virtual imaging data.

[0025] Determine whether the pixel density value is greater than or equal to the pixel density threshold. If the pixel density value is less than the pixel density threshold, set the feature recognizability score to zero.

[0026] If the pixel density value is greater than or equal to the pixel density threshold, then it is determined whether the sum of the proportion of the highlight overflow area and the proportion of the shadow occlusion area is less than or equal to the upper limit of the occlusion ratio. If the sum of the proportion of the highlight overflow area and the proportion of the shadow occlusion area is less than or equal to the upper limit of the occlusion ratio, then the ratio of the pixel density value to the preset pixel density benchmark value is used as the positive base score, and the positive base score is marked as the feature recognizability score.

[0027] If the sum of the proportion of the highlight overflow area and the proportion of the shadow occlusion area is greater than the upper limit of the occlusion ratio, the ratio of the pixel density value to the preset pixel density benchmark value is used as the positive base score, and the product of the sum of the proportion of the highlight overflow area and the proportion of the shadow occlusion area and the preset penalty coefficient is used as the penalty score. The difference is obtained by subtracting the penalty score from the positive base score. If the difference is less than zero, it is taken as zero. If the difference is not less than zero, the difference is marked as the feature identifiability score.

[0028] Furthermore, the process of constructing the objective function based on feature identifiability scores and deployment cost parameters includes:

[0029] Extract the feature identifiability scores of all target devices covered by each initial monitoring node, obtain the preset importance weights corresponding to each target device, multiply the feature identifiability scores of each target device by the corresponding preset importance weights and sum them to obtain the total score of global inspection effectiveness;

[0030] Obtain the single-node hardware cost and single-node construction cost from the deployment cost parameters, and multiply the initial number of monitoring nodes by the sum of the single-node hardware cost and single-node construction cost to obtain the total global deployment cost.

[0031] With the optimization direction of maximizing the total score of global inspection effectiveness and minimizing the total cost of global deployment, a multi-objective fitness function is constructed and marked as the objective function.

[0032] Furthermore, the process of iteratively optimizing the spatial coordinates and attitude parameters of the initial monitoring node using a heuristic search algorithm includes:

[0033] The spatial coordinates and attitude parameters of the initial monitoring nodes are encoded into population individuals, and the population individuals are substituted into the objective function to calculate the fitness value;

[0034] A fitness threshold is preset, and it is determined whether the fitness value is greater than or equal to the fitness threshold. If the fitness value is greater than or equal to the fitness threshold, the spatial coordinates and attitude parameters corresponding to the individuals in the population are extracted to generate the target placement scheme.

[0035] If the fitness value is less than the fitness threshold, crossover and mutation operations are performed on the individuals in the population to generate a new generation of individuals. The fitness value is then recalculated by substituting the results into the objective function until the preset maximum number of iterations is reached. Finally, the spatial coordinates and attitude parameters corresponding to the maximum fitness value are extracted to generate the target placement scheme.

[0036] Furthermore, the cloud platform is also equipped with a blind spot analysis and feedback module;

[0037] The blind spot analysis and feedback module is used to obtain the target deployment plan and extract the total global inspection effectiveness score of the target deployment plan at each discrete time node;

[0038] A preset validity threshold is set to determine whether the total validity score of the global inspection is less than the validity threshold. If the total validity score of the global inspection is less than the validity threshold, the corresponding discrete time nodes are extracted and marked as inspection blind spot periods.

[0039] If the total score of global inspection effectiveness is greater than or equal to the effectiveness threshold, then the discrete time node is marked as an effective inspection period.

[0040] Collect all inspection blind spot time periods and effective inspection time periods to generate a spatiotemporal distribution map of inspection blind spots.

[0041] Furthermore, after generating the spatiotemporal distribution map of inspection blind spots, it also includes:

[0042] Extract target devices with a feature identifiability score of zero within the blind spot period of the inspection and mark the target devices as blind spot devices;

[0043] Extract the spatial geometric features of the blind zone equipment and feed them back to the multi-objective collaborative optimization module. The multi-objective collaborative optimization module adds local monitoring nodes within a preset distance range around the blind zone equipment and re-executes the iterative optimization operation to update the target point layout scheme.

[0044] The beneficial effects of this invention are as follows: By extracting the spatial geometric features and material reflection properties of the target equipment to construct a three-dimensional semantic real-scene model, and combining it with a ray tracing algorithm to generate virtual light and shadow distribution data at different times, a target function is constructed based on feature identifiability scores and deployment cost parameters. The spatial coordinates and attitude parameters are iteratively optimized through a heuristic search algorithm to output the target deployment scheme. This solves the problems of existing schemes failing to accurately reflect the behavior of light reflection and refraction, leading to blindness caused by strong light, and the waste of hardware and construction resources caused by blindly stacking monitoring nodes. It effectively achieves the optimal balance between high identification rate of substation inspection and total global deployment cost. Attached Figure Description

[0045] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0046] Figure 1 is a schematic diagram of the substation camera automated deployment optimization system based on three-dimensional real scene modeling according to the present invention.

[0047] Figure 2 is a flowchart of the calculation process for the initial monitoring node feature identifiability score of the present invention. Detailed Implementation

[0048] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0049] Example 1:

[0050] Please refer to Figure 1. The substation camera automated deployment optimization system based on 3D real scene modeling includes a cloud platform, which is configured with a data acquisition and modeling module, a task constraint setting module, a spatiotemporal light and shadow simulation module, an imaging evaluation module, and a multi-objective collaborative optimization module.

[0051] The data acquisition and modeling module is used to acquire three-dimensional point cloud data of the target substation area, perform semantic segmentation on the three-dimensional point cloud data, extract the spatial geometric features and material reflection properties of the target equipment based on the semantic segmentation results, and construct a three-dimensional semantic reality model based on the spatial geometric features and material reflection properties.

[0052] The task constraint setting module is used to obtain the inspection task type of the target device, and allocate the corresponding pixel density threshold, observation angle range and occlusion ratio upper limit according to the inspection task type to generate a task semantic constraint set.

[0053] The spatiotemporal light and shadow simulation module is used to acquire the latitude and longitude data and time period of the target substation area. Based on the latitude and longitude data and time period, a solar trajectory model is constructed. Combined with the three-dimensional semantic real scene model and the task semantic constraint set, virtual light and shadow distribution data for different time periods are generated through ray tracing algorithm.

[0054] The imaging evaluation module is used to set the initial monitoring node in the three-dimensional semantic reality model, generate virtual imaging data corresponding to the viewpoint of the initial monitoring node by combining virtual light and shadow distribution data, input the virtual imaging data into the image feature evaluation algorithm, and output the feature recognizability score of the initial monitoring node.

[0055] The multi-objective collaborative optimization module is used to obtain the feature identifiability score and deployment cost parameters of the initial monitoring nodes. Based on the feature identifiability score and deployment cost parameters, an objective function is constructed. The spatial coordinates and attitude parameters of the initial monitoring nodes are iteratively optimized through a heuristic search algorithm to output the target deployment scheme.

[0056] This embodiment provides an automated substation camera deployment optimization system based on 3D real-scene modeling. The system acquires the original 3D point cloud data of the target substation area through lidar and oblique photography drones configured on site. The data acquisition and modeling module performs semantic segmentation on the acquired 3D point cloud data, extracting not only the spatial geometric features of target equipment such as the main transformer, but also the material reflection properties of its insulators and other components, and constructs a 3D semantic real-scene model to establish a digital twin containing physical attributes.

[0057] The task constraint setting module obtains the inspection task type of the target equipment through the production management system. Based on the inspection task type, it queries the preset task constraint mapping table and automatically assigns the corresponding pixel density threshold, observation angle range, and occlusion ratio upper limit. To clarify the specific task constraint standards, the task constraint mapping table presets specific numerical parameters: when the inspection task type is meter recognition, the system assigns a high pixel density threshold, specifically set to 120px / m, to meet the dial scale recognition standard, and a strict occlusion ratio upper limit, specifically set to 5%, which requires that more than 95% of the area is free from light and shadow interference; when the inspection task type is appearance defect detection, a medium pixel density threshold, specifically set to 60px / m, and a relatively broad occlusion ratio upper limit, specifically set to 15%, are assigned, thereby generating a task semantic constraint set, transforming the fuzzy inspection requirements into quantifiable mathematical constraints.

[0058] The spatiotemporal light and shadow simulation module acquires the latitude and longitude data of the target substation area through the positioning system and sets the time period. Based on the latitude and longitude data and the time period, it constructs a solar trajectory model. Combining the three-dimensional semantic real scene model and the task semantic constraint set, it uses a ray tracing algorithm to calculate the physical effects of the interaction between light and equipment materials at different time points, generating virtual light and shadow distribution data for different time periods. On this basis, the imaging evaluation module sets the initial monitoring node in the three-dimensional semantic real scene model, generates virtual imaging data of the corresponding viewpoint of the initial monitoring node by combining the virtual light and shadow distribution data, and inputs the virtual imaging data into the image feature evaluation algorithm to quantitatively output the feature identifiability score of the initial monitoring node.

[0059] Optionally, the multi-objective collaborative optimization module obtains the feature identifiability score and deployment cost parameters of the initial monitoring node through the input end, constructs an objective function based on the feature identifiability score and deployment cost parameters, iteratively optimizes the spatial coordinates and attitude parameters of the initial monitoring node through a heuristic search algorithm, and outputs the target deployment scheme.

[0060] The substation camera automated deployment optimization system based on 3D real-scene modeling provided in this embodiment can predict and avoid invalid deployment caused by strong light blindness in the all-weather monitoring scenario of substation by introducing spatiotemporal light and shadow simulation and task semantic constraints. This ensures that the generated deployment scheme meets the high-precision recognition requirements at the optical and semantic levels, and improves the effectiveness of substation automated inspection.

[0061] Optionally, the process of constructing a 3D semantic reality model includes:

[0062] Acquire 3D point cloud data of various target devices within the target substation area, and use clustering algorithms to separate the 3D point cloud data into independent device point cloud sets;

[0063] Extract the geometric topology of the independent device point cloud, match the geometric topology with a preset device component library containing the mapping relationship between device geometric topology and component semantic tags, and semantically annotate the independent device point cloud based on the matching results to obtain component semantic tags;

[0064] Obtain the surface material parameters corresponding to the semantic tags of the components, map the surface material parameters and component semantic tags to the geometric topology, and generate a 3D semantic reality model.

[0065] This embodiment further elaborates on the process of constructing a three-dimensional semantic reality model; three-dimensional point cloud data of various target devices in the target substation area are obtained through high-precision scanning equipment, and density-based clustering algorithms, such as noisy density-based spatial clustering algorithms or object sorting algorithms that identify cluster structures, are used to separate the three-dimensional point cloud data into independent sets of device point clouds.

[0066] The geometric topology of the independent device point cloud is extracted. Specifically, a 3D feature descriptor, such as a fast point feature histogram, is calculated for the independent device point cloud. The feature descriptor is then compared with the feature descriptor of the standard model in the preset device component library by calculating Euclidean distance or cosine similarity to select candidate models. Fine-grained registration is then performed using the iterative nearest point algorithm. In this process, the maximum number of iterations and the mean square error convergence threshold are set as the stopping conditions for the iterative nearest point algorithm. This allows the geometric topology to be matched with the preset device component library, which contains the mapping relationship between the geometric topology of the device and the semantic label of the component. Fine-grained semantic annotation is then performed on the independent device point cloud to obtain the semantic label of the component.

[0067] Based on this, the corresponding surface material parameters are obtained by querying the pre-set material database according to the semantic tags of the components. The surface material parameters include diffuse reflection coefficient, specular reflection coefficient, gloss index, transmission coefficient and refractive gloss index. The surface material parameters and component semantic tags are accurately mapped to the geometric topology to generate a three-dimensional semantic reality model containing geometric and physical properties.

[0068] In this embodiment, the process of constructing a three-dimensional semantic reality model, in the scenario of digital modeling of substation equipment, deeply binds geometric topology with semantic tags and material parameters, so that the model can realistically reflect the reflection and refraction behavior of light on different material surfaces, providing a physical basis for subsequent light and shadow simulation.

[0069] Optionally, the process of generating virtual light and shadow distribution data for different time periods using ray tracing algorithms includes:

[0070] Extract several discrete time nodes within the time period, and obtain the solar incident ray vector corresponding to each discrete time node based on the solar trajectory model;

[0071] The solar incident ray vector is projected onto the 3D semantic reality model, and the intersection point between the solar incident ray vector and the surface of the target device in the 3D semantic reality model is extracted.

[0072] The reflected ray vector and refracted ray vector are calculated based on the material reflection properties at the intersection point. The light intensity value of the target device surface is obtained based on the reflected ray vector, refracted ray vector and solar incident ray vector. The light intensity values ​​of each discrete time node are collected to generate virtual light and shadow distribution data.

[0073] This embodiment further explains the process of generating virtual light and shadow distribution data for different time periods using a ray tracing algorithm; it extracts several discrete time nodes within a time period, obtains the solar incident ray vector and the initial solar illumination intensity corresponding to each discrete time node based on the solar trajectory model; it projects the solar incident ray vector onto a three-dimensional semantic reality model, and uses spatial partitioning algorithms such as constructing hierarchical bounding boxes or octrees to accelerate calculations, and extracts the intersection points between the solar incident ray vector and the surface of the target device in the three-dimensional semantic reality model;

[0074] The reflected ray vector and refracted ray vector are calculated based on the material reflection properties at the intersection point. Specifically, this is based on the solar incident ray vector. With surface normal vector Calculate and obtain the reflected ray vector The calculation formula is:

[0075]

[0076] Regarding the vector of the refracted ray Introducing relative refractive index , which is the ratio of the refractive index of the incident medium to the refractive index of the target material, where the incident medium specifically refers to the air environment where the substation is located, and its refractive index is set to the standard atmospheric refractive index of 1.0. It is calculated using the following vector refraction formula:

[0077]

[0078] It is necessary to ensure that the value inside the square root is non-negative; otherwise, total internal reflection will occur. Zero vector; symbol This step represents the dot product operation of vectors; it obtains the light intensity value of the target device surface based on the reflected ray vector, refracted ray vector, and solar incident ray vector; in order to accurately assess the risk of blindness caused by strong light and avoid the loss of high light characteristics due to the fixed observation direction vector, this step adopts the maximum light intensity potential calculation method;

[0079] This method is based on the worst-case scenario principle, assuming that for any surface point, there exists a potential observation direction vector V and a reflected ray vector. Overlap, that is At this point, the specular reflection intensity reaches its theoretical maximum value; based on this assumption, virtual light and shadow distribution data independent of a specific viewpoint are calculated to expose all potential highlight clipping areas. The calculation formula is adjusted as follows:

[0080]

[0081] in, This represents the maximum potential illuminance at the intersection point, expressed in lx. The initial solar illumination intensity, expressed in lx; The diffuse reflectance coefficient is dimensionless. This indicates the operation of retrieving the maximum value of the parameter within the parentheses; The specular reflection coefficient is dimensionless; the standard illumination model includes the observation direction vector. and specular reflectance index However, under the assumption of maximum light intensity potential, since the line of sight is always assumed to coincide with the reflected light, Here, we directly take the coefficient. As the maximum contribution value for specular reflection, the parameter is thus included in the main formula. and Simplify and eliminate; The transmittance coefficient is dimensionless. The refractive gloss index is dimensionless; virtual light and shadow distribution data is generated by collecting the light intensity values ​​at various discrete time points.

[0082] This embodiment uses a ray tracing algorithm to generate virtual light and shadow distribution data for different time periods. In the dynamic environment simulation scenario of a substation, by introducing a maximum light intensity potential calculation model, the high light loss problem in traditional static mapping is effectively avoided, ensuring that the generated distribution data can cover the worst lighting conditions. This allows for early avoidance of optical blind spots during the deployment phase, demonstrating good temporal robustness.

[0083] Optionally, the process of generating virtual imaging data corresponding to the viewpoint of the initial monitoring node includes:

[0084] In the three-dimensional continuous space of the three-dimensional semantic reality model, the spatial coordinates and attitude parameters of several initial monitoring nodes are randomly generated;

[0085] Starting from spatial coordinates, a view frustum bounding box is constructed by combining attitude parameters, and the surface data of the 3D semantic reality model within the view frustum bounding box is extracted;

[0086] The surface data of the 3D semantic reality model within the view frustum bounding box is fused and rendered with the virtual light and shadow distribution data to generate a virtual 2D pixel matrix corresponding to the viewpoint, and the virtual 2D pixel matrix is ​​marked as virtual imaging data.

[0087] This embodiment further explains the process of generating virtual imaging data corresponding to the viewpoint of the initial monitoring node; the global bounding box of the three-dimensional semantic real scene model is extracted as an effective spatial sampling boundary. In the three-dimensional continuous space defined by the spatial sampling boundary, the spatial coordinates of several initial monitoring nodes are generated by a uniform random sampling method, and the corresponding attitude parameters are randomly generated in the three-dimensional rotation group space to avoid generating invalid nodes in an infinite space without boundaries.

[0088] Starting from spatial coordinates, a frustum bounding box is constructed by combining attitude parameters and preset camera intrinsic parameters. To ensure that the geometric boundary of the frustum in three-dimensional space is clear and conforms to the physical characteristics of real monitoring equipment, the preset camera intrinsic parameters adopt the standard fixed parameters of a typical industrial-grade substation inspection PTZ camera, specifically including: the preset focal length parameter is set to 8mm, the sensor size is set to 1 / 2.8 inch, corresponding to a physical size of approximately 5.27mm × 3.96mm, and the horizontal field of view calculated based on the above focal length and sensor size is approximately 36.4 degrees. At the same time, the near clipping plane is set to 0.5m and the far clipping plane is set to 150m. The system uses the above specific intrinsic parameter values ​​to calculate the boundary plane of the frustum, accurately construct the frustum bounding box, and extract the surface data of the three-dimensional semantic real scene model within the frustum bounding box.

[0089] The surface data of the 3D semantic real scene model within the view frustum bounding box is fused and rendered with the virtual light and shadow distribution data. Rasterization rendering technology is used to generate a virtual 2D pixel matrix and a corresponding component semantic mask map for the corresponding viewpoint. The virtual 2D pixel matrix and the component semantic mask map are jointly labeled as virtual imaging data.

[0090] Specifically, when generating the virtual two-dimensional pixel matrix, the maximum potential illumination intensity value at the intersection point is obtained using the following normalized mapping formula. Convert to grayscale value of pixels :

[0091]

[0092] in, This indicates the operation of retrieving the minimum value of the parameter within the parentheses. The preset maximum reference light intensity threshold for the environment. and These represent the x and y coordinates of a pixel in a virtual two-dimensional pixel matrix, respectively. The coordinates in the virtual two-dimensional pixel matrix are The maximum potential light intensity value corresponding to the pixel;

[0093] This embodiment generates virtual imaging data corresponding to the initial monitoring node's perspective. In the substation camera perspective pre-simulation scenario, it simulates the real camera imaging process in digital space, enabling the system to obtain image data in advance without actually installing hardware, providing direct data support for subsequent image quality assessment.

[0094] Optionally, the process of outputting the feature identifiability score of the initial monitoring node includes:

[0095] The virtual imaging data is input into the image feature evaluation algorithm, and the image feature evaluation algorithm outputs the pixel density value of the target device, the proportion of highlight overflow area, and the proportion of shadow occlusion area in the virtual imaging data.

[0096] Determine whether the pixel density value is greater than or equal to the pixel density threshold. If the pixel density value is less than the pixel density threshold, set the feature recognizability score to zero.

[0097] If the pixel density value is greater than or equal to the pixel density threshold, then it is determined whether the sum of the proportion of the highlight overflow area and the proportion of the shadow occlusion area is less than or equal to the upper limit of the occlusion ratio. If the sum of the proportion of the highlight overflow area and the proportion of the shadow occlusion area is less than or equal to the upper limit of the occlusion ratio, then the ratio of the pixel density value to the preset pixel density benchmark value is used as the positive base score, and the positive base score is marked as the feature recognizability score.

[0098] If the sum of the proportion of the highlight overflow area and the proportion of the shadow occlusion area is greater than the upper limit of the occlusion ratio, the ratio of the pixel density value to the preset pixel density benchmark value is used as the positive base score, and the product of the sum of the proportion of the highlight overflow area and the proportion of the shadow occlusion area and the preset penalty coefficient is used as the penalty score. The difference is obtained by subtracting the penalty score from the positive base score. If the difference is less than zero, it is taken as zero. If the difference is not less than zero, the difference is marked as the feature identifiability score.

[0099] This embodiment further elaborates on the process of scoring the feature identifiability of the output initial monitoring node, as shown in Figure 2; the virtual imaging data is input into the image feature evaluation algorithm; the model uses the component semantic tags of the target device to extract the target pixel region corresponding to the target device from the virtual imaging data; the total number of pixels in the target pixel region is divided by the actual physical surface area of ​​the target device in the three-dimensional semantic reality model, obtained by accumulating the area of ​​the mesh patches, and the square root of the division result is performed to calculate the pixel density value;

[0100] Simultaneously, the target pixel region is converted into a grayscale image. By setting a preset highlight grayscale threshold of 240 and a preset shadow grayscale threshold of 20, the grayscale image is segmented by double thresholds. The number of highlight pixels with grayscale values ​​greater than or equal to the highlight grayscale threshold and the number of shadow pixels with grayscale values ​​less than or equal to the shadow grayscale threshold are counted respectively. The number of highlight pixels and the number of shadow pixels are divided by the total number of pixels in the target pixel region to calculate the proportion of highlight overflow area and the proportion of shadow occlusion area.

[0101] When the pixel density value is less than the pixel density threshold, it indicates that the image clarity is insufficient to support the algorithm's recognition, and the system sets the feature recognizability score to zero; when the pixel density value is greater than or equal to the pixel density threshold, the system further judges the light and shadow occlusion situation.

[0102] When the sum of the proportion of highlight overflow area and the proportion of shadow occlusion area is less than or equal to the upper limit of occlusion ratio, the system uses the ratio of pixel density value to preset pixel density benchmark value as the positive base score. To prevent single-point score overflow due to excessively high pixel density of individual devices, which would affect the accuracy of global normalization logic, this embodiment sets the upper limit of the positive base score to 1.0. That is, when the calculated ratio is greater than 1.0, it is forcibly set to 1.0, and the processed positive base score is marked as the feature recognizability score. Among them, the preset pixel density benchmark value... The specific acquisition method is as follows: collect 5,000 equipment images that have been manually marked as clear and readable in the historical database of substations, calculate their average pixel density, and take the value as 150px / m;

[0103] When the sum of the proportion of highlight overflow area and the proportion of shadow occlusion area exceeds the upper limit of the occlusion ratio, a penalty mechanism is introduced. The ratio of the pixel density value to the preset pixel density baseline value is used as a positive base score, and the sum of the proportion of highlight overflow area and the proportion of shadow occlusion area is multiplied by the preset penalty coefficient. The product of the two is used as the penalty score. The positive base score, which is also subject to the upper limit of 1.0, is subtracted from the penalty score to obtain the difference. If the difference is less than zero, the feature identifiability score is zero. If the difference is not less than zero, the difference is marked as the feature identifiability score.

[0104] It should be noted that the preset penalty coefficient λ was obtained through the following gradient descent fitting process: a sample set of 1000 images with different highlight / shadow ratios, covering highlight-dominated, shadow-dominated, and mixed occlusion scenarios, was constructed, and experts were invited to score the recognizability of each sample as the label value. Define the loss function:

[0105]

[0106] in, This indicates that the summation operation is performed on all calculation results within the sample set; To substitute the current Calculated score; initialization Set the learning rate Iterate and update until convergence;

[0107]

[0108] in, This indicates an assignment or iterative update operation, which assigns the calculation result on the right to the variable on the left.

[0109] In calculating partial derivatives When, a subgradient strategy is adopted: when When the sample falls into the cutoff region, its contribution to the gradient is set as... Only by using the method to predict the score The effective activation samples are used for gradient backpropagation and accumulation; finally, the gradient is determined. The coefficient is 2.5; this fitting process ensures that even when using the summation form, the coefficients are consistent. It can also statistically fit the negative impact of light and shadow on image quality to the greatest extent possible. The calculation formula is as follows:

[0110]

[0111] in, The feature identifiability score is dimensionless and ranges from [0,1]. This represents the pixel density value, in pixels per square meter (px / m). The preset pixel density baseline value is obtained by statistically analyzing the average pixel density of clear historical inspection images of substations, specifically 150, with the unit being px / m; The preset penalty coefficient is obtained by fitting the light and shadow degradation curve using the gradient descent method. Specifically, it is 2.5, which is dimensionless. The percentage of the area of ​​highlight overflow is dimensionless. The percentage of the area occluded by shadow is dimensionless. The upper limit of the occlusion ratio is dimensionless;

[0112] This embodiment outputs the feature identifiability score of the initial monitoring node. In the substation monitoring screen quality assessment scenario, an assessment model is constructed through clear image processing logic, and reasonable boundary constraints are imposed on the score to ensure the objectivity and numerical stability of the scoring mechanism.

[0113] Optionally, the process of constructing the objective function based on feature identifiability scores and deployment cost parameters includes:

[0114] Extract the feature identifiability scores of all target devices covered by each initial monitoring node, obtain the preset importance weights corresponding to each target device, multiply the feature identifiability scores of each target device by the corresponding preset importance weights and sum them to obtain the total score of global inspection effectiveness;

[0115] Obtain the single-node hardware cost and single-node construction cost from the deployment cost parameters, and multiply the initial number of monitoring nodes by the sum of the single-node hardware cost and single-node construction cost to obtain the total global deployment cost.

[0116] With the optimization direction of maximizing the total score of global inspection effectiveness and minimizing the total cost of global deployment, a multi-objective fitness function is constructed and marked as the objective function.

[0117] This embodiment provides a detailed analysis of the process of constructing an objective function based on feature identifiability scores and deployment cost parameters; it extracts the feature identifiability scores of all target devices covered by each initial monitoring node, and obtains the preset importance weights corresponding to each target device from the configuration library. The feature identifiability scores of each target device are multiplied by their corresponding preset importance weights and then summed to obtain the total score of global inspection effectiveness.

[0118] The system input terminal obtains the single-node hardware cost and single-node construction cost parameters of the deployment cost. The initial number of monitoring nodes is multiplied by the sum of the single-node hardware cost and single-node construction cost to obtain the total global deployment cost. The optimization direction is to maximize the total global inspection effectiveness score and minimize the total global deployment cost. In order to eliminate the difference in dimensions and prevent a certain objective from dominating the optimization process, a preset global maximum deployment budget is introduced to normalize the cost, and the theoretical maximum effectiveness score is introduced to normalize the inspection effectiveness.

[0119] Before normalization, a zero-prevention safety check is performed on the normalization denominator. When the theoretical maximum effectiveness total score is zero, the corresponding effectiveness normalization term is set to zero. When the preset global maximum deployment budget is zero, the cost is considered to be seriously violated, and the multi-objective fitness function value is directly output as zero. After passing the zero-prevention safety check, a multi-objective fitness function is constructed and marked as the objective function. The calculation formula is as follows:

[0120]

[0121] in, The value of the multi-objective fitness function is dimensionless. The normalized weighting coefficient for inspection effectiveness was predetermined by combining the analytic hierarchy process with expert scoring. Since the primary goal of substation inspection is to ensure equipment monitoring coverage, effectiveness is given higher priority in the optimization process. The specific value is set to 0.7, which is dimensionless. The overall score for the effectiveness of the inspection is dimensionless. The theoretical maximum effectiveness score is the sum of the preset importance weights of all target devices.

[0122] Based on the premise that the upper limit of the positive base score is set to 1.0 in the aforementioned feature identifiability scoring calculation steps, this parameter ensures that the validity score ratio is always less than or equal to 1, thereby achieving a strict mapping to the [0,1] interval; The normalized weighting coefficient for site selection cost satisfies The corresponding specific value is set to 0.3, which is dimensionless; The total cost of global deployment is expressed in yuan. The preset global maximum deployment budget is in yuan.

[0123] This embodiment constructs an objective function based on feature identifiability scoring and deployment cost parameters. In the scenario of economic analysis of substation deployment schemes, the recognition rate and cost are incorporated into a unified mathematical model. Through a dual normalization mechanism and boundary truncation processing, the numerical stability of the algorithm in multi-objective trade-offs is ensured, avoiding the waste of resources caused by blindly stacking cameras, and achieving a balance between efficiency and cost.

[0124] Example 2:

[0125] The process of iteratively optimizing the spatial coordinates and attitude parameters of the initial monitoring node using a heuristic search algorithm includes:

[0126] The spatial coordinates and attitude parameters of the initial monitoring nodes are encoded into population individuals, and the population individuals are substituted into the objective function to calculate the fitness value;

[0127] A fitness threshold is preset, and it is determined whether the fitness value is greater than or equal to the fitness threshold. If the fitness value is greater than or equal to the fitness threshold, the spatial coordinates and attitude parameters corresponding to the individuals in the population are extracted to generate the target placement scheme.

[0128] If the fitness value is less than the fitness threshold, crossover and mutation operations are performed on the individuals in the population to generate a new generation of individuals. The fitness value is then recalculated by substituting the results into the objective function until the preset maximum number of iterations is reached. Finally, the spatial coordinates and attitude parameters corresponding to the maximum fitness value are extracted to generate the target placement scheme.

[0129] This embodiment further specifies the process of iteratively optimizing the spatial coordinates and attitude parameters of the initial monitoring node using a heuristic search algorithm. The heuristic search algorithm specifically adopts a genetic algorithm. The spatial coordinates and attitude parameters of the initial monitoring node are encoded into population individuals in the form of real number strings, and the population individuals are substituted into the objective function to calculate the fitness value of each individual.

[0130] The system presets a fitness threshold, which is determined as follows: 100 successful substation layout schemes implemented and accepted within the past year are extracted, and the average fitness value of these schemes under the current objective function is calculated. For example, 0.92, and a safety redundancy factor is introduced. Specifically, it is set to 0.05 to address environmental uncertainty, and then the final fitness threshold is calculated. The calculation formula is:

[0131]

[0132] For example, 0.92 * 0.95 = 0.874, to ensure that the output scheme meets the engineering baseline requirements; when the fitness value is greater than or equal to the fitness threshold, extract the spatial coordinates and attitude parameters corresponding to the individuals in the population, and directly generate the target placement scheme;

[0133] Based on this, when the fitness value is less than the fitness threshold, the system performs crossover and mutation operations on the population individuals to generate a new generation of population individuals. Specifically, a binary tournament selection strategy is used to select the best parent individuals by eliminating the inferior individuals in the population based on their fitness values.

[0134] During this process, the system presets the crossover probability and mutation probability to obtain the initial total number of monitoring nodes. Since each node contains 3 spatial coordinate parameters and 3 pose parameters, the dimension of individual parameters in the population is determined to be... The mutation probability is set to the reciprocal of the parameter dimension, and a random number between 0 and 1 is generated for the selected parent individual:

[0135] When the random number is less than or equal to the crossover probability, the spatial coordinates and attitude parameters of the selected parent individual are recombined by simulating the binary crossover operator, and a new random number is generated after the recombination.

[0136] When the new random number is less than or equal to the mutation probability, the parameters after recombination are fine-tuned by the polynomial mutation operator to maintain the diversity of the population. The fitness value is recalculated by substituting it into the objective function. The process is repeated until the preset maximum number of iterations is reached. The spatial coordinates and attitude parameters corresponding to the maximum fitness value are extracted to generate the final target placement scheme.

[0137] This embodiment uses a heuristic search algorithm to iteratively optimize the spatial coordinates and attitude parameters of the initial monitoring nodes. In the scenario of searching for massive deployment parameters in a substation, the algorithm utilizes the global optimization characteristics of the genetic algorithm to escape local optima. Combined with specific threshold calculation logic and a strict probability triggering mechanism, it quickly converges to the optimal camera layout parameters that balance coverage and image quality, demonstrating the algorithm's optimization robustness.

[0138] Example 3:

[0139] The cloud also includes a blind spot analysis and feedback module;

[0140] The blind spot analysis and feedback module is used to obtain the target deployment plan and extract the total global inspection effectiveness score of the target deployment plan at each discrete time node;

[0141] A preset validity threshold is set to determine whether the total validity score of the global inspection is less than the validity threshold. If the total validity score of the global inspection is less than the validity threshold, the corresponding discrete time nodes are extracted and marked as inspection blind spot periods.

[0142] If the total score of global inspection effectiveness is greater than or equal to the effectiveness threshold, then the discrete time nodes are marked as effective inspection periods.

[0143] Collect all inspection blind spot time periods and effective inspection time periods to generate a spatiotemporal distribution map of inspection blind spots.

[0144] This embodiment provides a detailed explanation of the operation mechanism of the blind spot analysis and feedback module for cloud communication connections; the blind spot analysis and feedback module obtains the target deployment scheme output by the multi-target collaborative optimization module and extracts the total global inspection effectiveness score of the target deployment scheme at each discrete time node;

[0145] The system presets an effectiveness threshold, which is determined in advance based on the product of the theoretical maximum total effectiveness score and the preset minimum acceptable effectiveness ratio coefficient, to ensure that the dimension of the threshold is consistent with the total effectiveness score of the global inspection.

[0146] When the total score of global inspection effectiveness is less than the effectiveness threshold, it indicates that the lighting conditions at that time point seriously affected the overall inspection quality. The system extracts the corresponding discrete time point and marks the discrete time point as the inspection blind spot period.

[0147] When the total score of global inspection effectiveness is greater than or equal to the effectiveness threshold, the system marks the discrete time node as an effective inspection period; the system collects all inspection blind spot periods and effective inspection periods to generate an intuitive spatiotemporal distribution map of inspection blind spots;

[0148] This embodiment generates a spatiotemporal distribution map of inspection blind spots through the blind spot analysis and feedback module. In the scenario of substation inspection plan scheduling, it not only provides the spatial location of the points, but also outputs inspection suggestions in the time dimension in combination with clear validity judgment criteria. This enables the operation and maintenance system to avoid issuing inspection tasks during periods of poor lighting, thereby improving the reliability of inspection data.

[0149] Optionally, after generating the spatiotemporal distribution map of inspection blind spots, the following steps are also included:

[0150] Extract target devices with a feature identifiability score of zero within the blind spot period of the inspection and mark them as blind spot devices.

[0151] Extract the spatial geometric features of the blind zone equipment and feed them back to the multi-objective collaborative optimization module. The multi-objective collaborative optimization module adds local monitoring nodes within a preset distance range around the blind zone equipment and re-executes the iterative optimization operation to update the target point layout scheme.

[0152] This embodiment further clarifies the closed-loop optimization steps after generating the spatiotemporal distribution map of the inspection blind zone; within the identified inspection blind zone time period, the system extracts the target equipment with a feature identifiability score of zero and marks the target equipment as a blind zone equipment; extracts the spatial geometric features of the blind zone equipment and feeds the spatial geometric features of the blind zone equipment as a compensation signal to the multi-target collaborative optimization module;

[0153] The multi-target collaborative optimization module forcibly adds local monitoring nodes within a preset distance range around the blind zone device. The preset distance range is determined by comprehensively considering the physical size bounding box of the blind zone device's spatial geometric features, the preset shortest focusing distance of the camera, and the maximum effective recognition distance. Specifically, with the geometric center of the physical size bounding box as the center of a sphere, the sum of the maximum circumscribed sphere radius of the physical size bounding box and the shortest focusing distance is used as the minimum generation radius, and the maximum effective recognition distance is used as the maximum generation radius. A spherical effective sampling space is constructed as the preset distance range to ensure that the newly added nodes have a good field of view and are neither too close to the target device to avoid focusing failure nor too far to avoid insufficient resolution. The parameters of the existing nodes are locked, and the iterative optimization operation is re-executed only for the newly added local monitoring nodes, updating and outputting the corrected target point layout scheme.

[0154] At the same time, the maximum number of blind patching iterations is set. Specifically set as Next; if the same blind zone device is executed If the feature identifiability score of a local monitoring node remains zero after the addition and iterative optimization, the blind spot patching operation will be stopped, an abnormal occlusion alarm will be output to prompt manual investigation, and the current blind spot patching loop will be exited.

[0155] This embodiment focuses on the subsequent optimization of the spatiotemporal distribution map of inspection blind spots. In the scenario of monitoring blind spot filling in complex obstructed areas of substations, a closed-loop optimization mechanism is formed to discover blind spots and fill them in a targeted manner. By eliminating hard-to-observe blind spot devices at fixed points, the system ensures the coverage of the final target deployment scheme without blind spots and further enhances the system's environmental adaptability.

[0156] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A substation camera automated deployment optimization system based on 3D real-scene modeling, characterized in that, This includes a cloud platform, which is equipped with a data acquisition and modeling module, a task constraint setting module, a spatiotemporal light and shadow simulation module, an imaging evaluation module, and a multi-objective collaborative optimization module. The data acquisition and modeling module acquires 3D point cloud data of the target substation area, performs semantic segmentation on the 3D point cloud data, extracts the spatial geometric features and material reflection properties of the target equipment based on the semantic segmentation results, and constructs a 3D semantic reality model based on the spatial geometric features and material reflection properties. The task constraint setting module acquires the inspection task type of the target equipment, assigns corresponding pixel density thresholds, observation angle ranges, and occlusion ratio upper limits according to the inspection task type, and generates a task semantic constraint set. The spatiotemporal light and shadow simulation module acquires the latitude and longitude data and time period of the target substation area, constructs a solar trajectory model based on the latitude and longitude data and time period, and combines it with the 3D semantic reality model. The system consists of a model and a task semantic constraint set, which generates virtual light and shadow distribution data for different time periods using a ray tracing algorithm. The imaging evaluation module sets initial monitoring nodes in the 3D semantic reality model, generates virtual imaging data corresponding to the viewpoint of the initial monitoring node by combining the virtual light and shadow distribution data, inputs the virtual imaging data into the image feature evaluation algorithm, and outputs the feature recognizability score of the initial monitoring node. The multi-objective collaborative optimization module obtains the feature recognizability score and placement cost parameters of the initial monitoring node, constructs an objective function based on the feature recognizability score and placement cost parameters, iteratively optimizes the spatial coordinates and pose parameters of the initial monitoring node through a heuristic search algorithm, and outputs the target placement scheme. The process of outputting the feature identifiability score of the initial monitoring node includes: inputting virtual imaging data into an image feature evaluation algorithm; the image feature evaluation algorithm outputs the pixel density value, highlight overflow area ratio, and shadow occlusion area ratio of the target device in the virtual imaging data; determining whether the pixel density value is greater than or equal to a pixel density threshold; if the pixel density value is less than the pixel density threshold, the feature identifiability score is set to zero; if the pixel density value is greater than or equal to the pixel density threshold, determining whether the sum of the highlight overflow area ratio and the shadow occlusion area ratio is less than or equal to the upper limit of the occlusion ratio; if the sum of the highlight overflow area ratio and the shadow occlusion area ratio is less than or equal to the upper limit of the occlusion ratio, the ratio of the pixel density value to a preset pixel density benchmark value is used as a positive base score, and the positive base score is marked as the feature identifiability score; if the sum of the highlight overflow area ratio and the shadow occlusion area ratio is greater than the upper limit of the occlusion ratio, the ratio of the pixel density value to a preset pixel density benchmark value is used as a positive base score, and the highlight overflow area ratio is set as the upper limit of the occlusion ratio. The product of the sum of the overflow area ratio and the shadow occlusion area ratio and the preset penalty coefficient is used as the penalty score. The difference is obtained by subtracting the penalty score from the positive base score. If the difference is less than zero, it is set to zero. If the difference is not less than zero, the difference is marked as the feature identifiability score. The process of constructing an objective function based on the feature identifiability score and the deployment cost parameters includes: extracting the feature identifiability scores of all target devices covered by each initial monitoring node, obtaining the preset importance weights corresponding to each target device, multiplying the feature identifiability scores of each target device by the corresponding preset importance weights and summing the results to obtain the total global inspection effectiveness score; obtaining the single-node hardware cost and single-node construction cost in the deployment cost parameters, multiplying the number of initial monitoring nodes by the sum of the single-node hardware cost and single-node construction cost to obtain the total global deployment cost; constructing a multi-objective fitness function with the optimization direction of maximizing the total global inspection effectiveness score and minimizing the total global deployment cost, and marking the multi-objective fitness function as the objective function.

2. The substation camera automated deployment optimization system based on 3D real-scene modeling according to claim 1, characterized in that, The process of constructing a 3D semantic reality model includes: acquiring 3D point cloud data of various target devices within the target substation area; separating the 3D point cloud data into targets using a clustering algorithm to obtain independent device point cloud sets; extracting the geometric topology of the independent device point cloud sets; matching the geometric topology with a pre-defined equipment component library containing the mapping relationship between equipment geometric topology and component semantic labels; semantically labeling the independent device point cloud sets based on the matching results to obtain component semantic labels; obtaining the surface material parameters corresponding to the component semantic labels; mapping the surface material parameters and component semantic labels to the geometric topology to generate a 3D semantic reality model.

3. The substation camera automated deployment optimization system based on 3D real-scene modeling according to claim 1, characterized in that, The process of generating virtual light and shadow distribution data for different time periods using ray tracing algorithms includes: extracting several discrete time nodes within a time period; obtaining the solar incident ray vector corresponding to each discrete time node based on the solar trajectory model; projecting the solar incident ray vector onto a 3D semantic reality model; extracting the intersection points between the solar incident ray vector and the surface of the target device in the 3D semantic reality model; calculating the reflected ray vector and the refracted ray vector based on the material reflection properties at the intersection points; obtaining the illumination intensity value of the target device surface based on the reflected ray vector, the refracted ray vector, and the solar incident ray vector; and collecting the illumination intensity values ​​of each discrete time node to generate virtual light and shadow distribution data.

4. The substation camera automated deployment optimization system based on 3D real-scene modeling according to claim 1, characterized in that, The process of generating virtual imaging data corresponding to the initial monitoring node's viewpoint includes: randomly generating the spatial coordinates and attitude parameters of several initial monitoring nodes in the three-dimensional continuous space of the three-dimensional semantic reality model; constructing a view frustum bounding box based on the spatial coordinates and attitude parameters, and extracting the surface data of the three-dimensional semantic reality model within the view frustum bounding box; fusing and rendering the surface data of the three-dimensional semantic reality model within the view frustum bounding box with virtual light and shadow distribution data to generate a virtual two-dimensional pixel matrix corresponding to the viewpoint, and marking the virtual two-dimensional pixel matrix as virtual imaging data.

5. The substation camera automated deployment optimization system based on 3D real-scene modeling according to claim 1, characterized in that, The process of iteratively optimizing the spatial coordinates and attitude parameters of the initial monitoring nodes using a heuristic search algorithm includes: encoding the spatial coordinates and attitude parameters of the initial monitoring nodes into population individuals; substituting the population individuals into the objective function to calculate their fitness values; setting a fitness threshold and determining whether the fitness value is greater than or equal to the fitness threshold; if the fitness value is greater than or equal to the fitness threshold, extracting the spatial coordinates and attitude parameters corresponding to the population individuals to generate a target placement scheme; if the fitness value is less than the fitness threshold, performing crossover and mutation operations on the population individuals to generate a new generation of population individuals, resubmitting them into the objective function to calculate their fitness values, until the preset maximum number of iterations is reached, extracting the spatial coordinates and attitude parameters corresponding to the maximum fitness value to generate the target placement scheme.

6. The substation camera automated deployment optimization system based on 3D real-scene modeling according to claim 5, characterized in that, The cloud-based system is also equipped with a blind spot analysis and feedback module. The blind spot analysis and feedback module is used to obtain the target deployment plan and extract the total global inspection effectiveness score of the target deployment plan at each discrete time node. A preset validity threshold is set to determine whether the total validity score of the global inspection is less than the validity threshold. If the total validity score of the global inspection is less than the validity threshold, the corresponding discrete time nodes are extracted and marked as inspection blind spot periods. If the total score of global inspection effectiveness is greater than or equal to the effectiveness threshold, then the discrete time node is marked as an effective inspection period. Collect all inspection blind spot time periods and effective inspection time periods to generate a spatiotemporal distribution map of inspection blind spots.

7. The substation camera automated deployment optimization system based on 3D real-scene modeling according to claim 6, characterized in that, After generating the spatiotemporal distribution map of the inspection blind zone, the process also includes: extracting target devices with a feature identifiability score of zero within the inspection blind zone period and marking the target devices as blind zone devices; extracting the spatial geometric features of the blind zone devices and feeding the spatial geometric features of the blind zone devices back to the multi-target collaborative optimization module; the multi-target collaborative optimization module adds local monitoring nodes within a preset distance range around the blind zone devices and re-executes iterative optimization operations to update the target deployment scheme.

Images