Intelligent classification system for outdoor substation equipment in combination with point cloud data recognition

Abstract

The application provides an outdoor transformer substation equipment intelligent classification system combined with point cloud data recognition, relates to the technical field of point cloud data processing, and comprises: an echo analysis module, which is used for establishing a three-dimensional point cloud data set; a ground model fitting module, which is used for generating a transformer substation point cloud data set; a density clustering analysis module, which is used for generating a first point cloud segmentation result; a segmentation boundary optimization module, which is used for generating a point cloud single segmentation result; and an equipment classification module, which is used for performing equipment classification. Through the application, the technical problems of low equipment recognition efficiency, multi-source point cloud registration error accumulation, and misclassification caused by fuzzy segmentation boundary due to the dependence on manual segmentation in traditional point cloud processing can be solved, the technical effects of optimizing point cloud collection accuracy based on multi-echo analysis, realizing high-robustness segmentation through iterative fitting and density clustering, improving boundary definition accuracy by fusing normal vectors and curvature analysis, and realizing full-automatic high-precision classification of transformer substation equipment can be achieved.

Description

Intelligent classification system for outdoor substation equipment based on point cloud data recognition Technical Field

[0001] This application relates to the field of point cloud data processing technology, specifically to an intelligent classification system for outdoor substation equipment that combines point cloud data recognition. Background Technology

[0002] With the rapid development of the power industry, substations, as a crucial component of power transmission and distribution networks, possess a wide variety and complexity of equipment. Traditional methods for managing and classifying substation equipment largely rely on manual identification and recording. This is not only labor-intensive and inefficient, but also prone to inaccuracies and delays in equipment management information due to the high error rate inherent in manual operations, potentially leading to safety hazards. Furthermore, on-site operations are often limited by external factors such as weather and geographical location, resulting in slow data collection and updates, making it difficult to reflect the real-time status of substation equipment in a timely manner.

[0003] With the development of 3D laser scanning technology, point cloud data acquisition has become an important means of obtaining high-precision equipment information. However, traditional point cloud data processing and equipment classification methods still have many shortcomings. Existing technologies usually rely only on simple acquisition and basic processing of point cloud data, lacking efficient subsequent intelligent analysis methods, resulting in equipment classification and management still requiring a large amount of manual intervention.

[0004] Therefore, there is an urgent need for an efficient and intelligent equipment classification and management system that can automatically identify and classify substation equipment from point cloud data, improve management efficiency and accuracy, and avoid the risks of manual operation. Summary of the Invention

[0005] The purpose of this application is to provide an intelligent classification system for outdoor substation equipment that combines point cloud data recognition, in order to solve the technical problems of low equipment recognition efficiency, accumulation of multi-source point cloud registration errors, and misclassification caused by blurred segmentation boundaries in traditional point cloud processing due to reliance on manual segmentation.

[0006] In view of the above problems, this application provides an intelligent classification system for outdoor substation equipment that combines point cloud data recognition.

[0007] This application provides an intelligent classification system for outdoor substation equipment based on point cloud data recognition. The system includes: an echo analysis module for performing omnidirectional scanning of the target outdoor substation using a 3D laser scanner and conducting multi-echo analysis of laser pulses to establish a 3D point cloud dataset after equipment overlap segmentation; a ground model fitting module for preprocessing the 3D point cloud dataset, iteratively fitting a ground point cloud model, deleting ground point cloud data based on the fitting results, and generating a substation site cloud dataset; a density clustering analysis module for performing density clustering segmentation on the substation site cloud dataset to generate a first point cloud segmentation result; a segmentation boundary optimization module for performing normal vector and curvature similarity analysis on the segmentation boundary of the first point cloud segmentation result, optimizing the segmentation boundary, and generating a point cloud unit segmentation result; and an equipment classification module for performing multi-view 2D image rendering on the point cloud unit segmentation result, and classifying the equipment based on the 2D image rendering result.

[0008] The technical solution provided in this application has at least the following technical effects or advantages:

[0009] The aforementioned intelligent classification system for outdoor substation equipment, which combines point cloud data recognition, includes an echo analysis module, a ground model fitting module, a density clustering analysis module, a segmentation boundary optimization module, and an equipment classification module. The echo analysis module uses a 3D laser scanner to perform a 3D all-around scan of the target substation and conducts multi-echo analysis of laser pulses to generate a processed 3D point cloud dataset. Based on this, the ground model fitting module preprocesses the point cloud data, uses an iterative method to fit a ground point cloud model, and deletes ground point cloud data, thereby generating an accurate substation point cloud dataset. Next, the density clustering analysis module performs density clustering segmentation on the substation point cloud dataset to obtain preliminary point cloud segmentation results. To optimize the segmentation results, the segmentation boundary optimization module performs boundary normal vector and curvature similarity analysis on these preliminary segmentation results, further refining the segmentation and generating more accurate point cloud unit segmentation results. Finally, the equipment classification module performs multi-view 2D image rendering based on the optimized point cloud unit segmentation results, and classifies the equipment based on the rendering results. The entire system, through automated point cloud data processing and intelligent classification, not only improves the efficiency and accuracy of substation equipment management, but also reduces manual intervention and ensures the real-time nature of data updates and equipment management.

[0010] The above description is merely an overview of the technical solution of this application. To better understand the technical means of this application and to facilitate its implementation according to the description, and to make the above and other objects, features, and advantages of this application more apparent, specific embodiments of this application are described below. It should be understood that the content described in this section is not intended to identify key or important features of the embodiments of this application, nor is it intended to limit the scope of this application. Other features of this application will become readily apparent through the following description. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely exemplary. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0012] Figure 1 is a schematic diagram of the structure of the intelligent classification system for outdoor substation equipment based on point cloud data recognition in this application;

[0013] Figure 2 is a flowchart of the echo analysis module of the intelligent classification system for outdoor substation equipment based on point cloud data recognition in this application.

[0014] Explanation of reference numerals in the attached figures:

[0015] Echo analysis module 11, ground model fitting module 12, density clustering analysis module 13, segmentation boundary optimization module 14, equipment classification module 15. Detailed Implementation

[0016] This application provides an intelligent classification system for outdoor substation equipment that combines point cloud data recognition. This system solves the technical problems of low equipment recognition efficiency, accumulation of multi-source point cloud registration errors, and misclassification caused by blurred segmentation boundaries in traditional point cloud processing due to reliance on manual segmentation. It achieves the technical effect of fully automatic and high-precision classification of substation equipment by optimizing point cloud acquisition accuracy based on multi-echo analysis, realizing highly robust segmentation through iterative fitting and density clustering, and improving the accuracy of boundary definition by fusing normal vectors and curvature analysis.

[0017] The technical solutions of this application will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. It should be understood that this application is not limited to the exemplary embodiments described herein. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application. It should also be noted that, for ease of description, only the parts related to this application are shown in the accompanying drawings, not all of them.

[0018] As illustrated in Figure 1, this application provides an intelligent classification system for outdoor substation equipment that combines point cloud data recognition. The system specifically includes the following modules:

[0019] The echo analysis module 11 is used to perform a full-range scan of the target outdoor substation using a 3D laser scanner, and to perform multi-echo analysis of the laser pulses to establish a 3D point cloud dataset after the equipment overlaps and is segmented.

[0020] Specifically, in the echo analysis module 11, during point cloud data acquisition at outdoor substations, issues such as dense equipment layout, obstructed views, and point cloud overlap often lead to the mixing of point clouds from multiple devices, forming indistinguishable overlapping areas. This phenomenon is usually caused by the scanning angle of LiDAR or cameras affecting the distribution of point clouds, especially when viewed from certain angles, where the point clouds of devices in front and behind may project onto the same location, creating a virtual overlap. To solve this problem, the system terminal utilizes LiDAR's multiple return technology to distinguish between devices in front and behind. Specifically, the system terminal performs omnidirectional scanning using multiple 3D laser scanners (such as the Leica RTC360 or FARO Focus S series). These 3D laser scanners are set up at multiple scanning stations within the substation to ensure coverage of all devices. Each laser pulse generated during a scan may encounter multiple reflecting surfaces, producing different echoes. Common echoes include the first return, intermediate returns, and last return. The first return typically originates from the surface of the object closest to the laser source, such as the equipment casing, cables, or supports. Intermediate returns may occur after the laser penetrates transparent or sparsely populated objects, such as leaves, glass covers, or cable nets. The last return usually originates from the farthest reflecting surface, which may be the metal frame behind the equipment or the ground. By analyzing these multi-echo data in conjunction with a preset spacing threshold, the system terminal can effectively distinguish the point clouds of the front and rear devices. For example, when a laser beam illuminates a transformer in front, some laser light may penetrate the transformer's gaps and reflect at the switchgear behind. In this case, the first return records the point cloud data of the transformer's surface, while the last return records the point cloud data of the switchgear. By calculating the spacing between the echoes, it is possible to determine whether there are gaps or transparent objects between them, thereby separating the point cloud data of different devices. Multi-echo analysis can accurately establish a 3D point cloud dataset after overlapping and segmenting of equipment, improving the clarity and accuracy of the point cloud data and providing a reliable foundation for subsequent equipment identification and classification.

[0021] Furthermore, as shown in Figure 2, the echo analysis module also includes:

[0022] The system receives laser pulse echo data from each scan performed by the 3D laser scanner; extracts multiple echo data from the laser pulse echo data; calculates the echo spacing between adjacent first and second echo data in the multiple echo data to generate a first echo spacing; determines whether the first echo spacing is greater than a preset spacing threshold; if so, performs preliminary point cloud overlap segmentation based on the first and second echo data to generate segmented point cloud data, and adds it to the 3D point cloud dataset.

[0023] In a preferred embodiment, when receiving laser pulse echo data from each scan by the 3D laser scanner, the system terminal extracts multiple echo data, including a first echo, a second echo, and subsequent echo data. These echo data correspond to the reflection information of the laser pulse at different depths or on different object surfaces, which can help distinguish between preceding and following devices and their overlapping areas. Subsequently, the system terminal processes these echo data, especially calculating the echo spacing between adjacent first and second echo data to obtain the first echo spacing. The first echo spacing refers to the distance between two echoes, which can reflect the presence of gaps or transparent objects between devices. It is obtained by calculating the time difference or distance difference between the first and second echoes. When the echo spacing between the first and second echoes is large, it indicates that there is a certain gap or transparent object between the devices corresponding to these two echoes, such as a void, cable network, glass cover, etc. This information can be used to effectively distinguish devices. To ensure accurate segmentation, a preset echo spacing threshold is established. This threshold determines whether two echoes belong to different devices or objects. It is based on the common layout of various devices in a substation and the typical spacing between them. For example, in a substation, the distance between a transformer and a switchgear is usually 3 meters, so the preset spacing threshold can be set to 3 meters. When the calculated echo spacing is less than or equal to the preset threshold, it indicates that the two echoes may originate from adjacent surfaces of the same device, and the system terminal will merge them. When the calculated echo spacing is greater than the preset threshold, it indicates that there is a significant gap or transparent object between the devices, and they can be considered to belong to different devices. At this point, the system terminal will perform preliminary point cloud overlap segmentation based on the data from the first and second echoes. That is, it will mark the point clouds corresponding to the first and second echoes as different devices or objects, thereby generating segmented point cloud data. This segmented point cloud data will be added to the 3D point cloud dataset as input for subsequent processing. It is important to note that if the point cloud data only contains the final merged point cloud without recording information from each echo, multi-echo analysis is impossible. To avoid this, echo analysis must be performed during real-time data acquisition during the scanning process. This ensures accurate processing of each echo data point and avoids the need for complex region segmentation operations based on the generated point cloud data later. This is not only difficult but can also lead to low processing efficiency and error accumulation. Therefore, the system terminal should perform echo analysis during scanning to accurately distinguish overlapping areas between devices and perform effective data processing and segmentation. Through this process, multi-echo information can be effectively utilized to solve the problem of mixed device point clouds, generating a high-quality 3D point cloud dataset, providing a reliable foundation for subsequent device segmentation and classification.

[0024] For example, a 3D laser scanner is used to scan an outdoor substation. During the scanning process, the laser pulse encounters the reflective surfaces of multiple devices, generating multiple echo data, as shown in Tables 1, 2, and 3 below.

[0025] Table 1: Laser Pulse Echo Data Table

[0026]

[0027] Table 2: Echo Spacing Calculation Table

[0028]

[0029]

[0030] Table 3: Laser Pulse Echo Data Table

[0031]

[0032]

[0033] Table 1 presents an example of laser pulse echo data, recording multiple echoes generated during each laser pulse scan, including echo type, echo time, echo distance, and corresponding reflective surface description. This data allows for preliminary identification of the point cloud distribution of different devices. Table 2 illustrates the calculation process for echo spacing. By calculating the distance between adjacent echoes and comparing it to a preset spacing threshold, it determines whether point cloud data segmentation is necessary. For example, when the echo spacing is greater than the threshold, the point cloud data of different devices is segmented into independent point cloud clusters; when the echo spacing is less than or equal to the threshold, the point cloud data is merged. Table 3 displays the final point cloud segmentation results. Based on the echo spacing determination, the point cloud data is segmented into different devices or objects, recording the point cloud source, distance range, and point cloud description for each device. These segmentation results provide a high-quality point cloud data foundation for subsequent device classification and management. These three tables collectively demonstrate how the system uses multi-echo analysis technology to solve the problem of mixed device point clouds and generate high-quality 3D point cloud datasets, supporting intelligent management of substation equipment.

[0034] Furthermore, this application provides a method for determining whether the first echo spacing is greater than a preset spacing threshold, including:

[0035] If the first echo spacing is less than or equal to the preset spacing threshold, the first echo data and the second echo data are merged, and the processed point cloud is added to the three-dimensional point cloud dataset.

[0036] In one optional implementation, when the echo spacing between the first and second echoes is less than or equal to a preset threshold, it indicates that there is no obvious gap or transparent object between them, and they may belong to different parts of the same device. In this case, the system terminal will not separate them, but will merge the two echo data. The merging process typically involves combining the point cloud data of both to form a continuous point cloud dataset, so that it can be analyzed as a whole in subsequent processing. During the merging process, duplicate points are removed to ensure higher quality point cloud data after merging. Finally, the merged point cloud data will be added to a 3D point cloud dataset to ensure that complete device information is accurately recorded and processed.

[0037] The ground model fitting module 12 is used to preprocess the three-dimensional point cloud dataset, perform iterative fitting of the ground point cloud model, delete the ground point cloud data according to the fitting results, and generate a substation site cloud dataset.

[0038] Specifically, in the ground model fitting module 12, after obtaining the 3D point cloud dataset, it needs to be preprocessed, including denoising and registration. Denoising is performed using the Statistical Outlier Removal (SOR) algorithm, which calculates the distance d between each point and its k nearest neighbors. k (p i ), and assume that the point cloud data follows a Gaussian distribution. For each point p i The mean μ and standard deviation σ of this point can be calculated using the following formula:

[0039] Where n is the total number of points in the 3D point cloud dataset. If the distance of a data point exceeds the mean plus a preset multiple of the standard deviation, the point is considered an outlier and needs to be removed. By removing outliers, the interference of noise on the point cloud data can be reduced, ensuring the accuracy of subsequent analysis. The alignment criterion is to use the Iterative Closest Point (ICP) algorithm to accurately align point cloud data from different scanning positions. The ICP algorithm performs registration by minimizing the Euclidean distance between corresponding points in the source and target point clouds, ensuring that point cloud data from different sources are accurately aligned in spatial position, forming a complete 3D point cloud dataset. Specifically, the ICP algorithm optimizes the following objective function by continuously adjusting the rotation matrix R and the translation vector t:

[0040] Where E(R,t) is the objective function, representing the error between the source point cloud and the target point cloud; N p It is the number of points in the source point cloud; x i The coordinates of point i in the target point cloud; p i These are the coordinates of point i in the source point cloud. Spatial alignment of the point cloud data is ensured through iterative optimization. After preprocessing, the system terminal enters the ground point cloud model fitting stage. The iterative fitting process uses a method based on the Random Sample Consensus (RANSAC) algorithm. During this process, the preprocessed point cloud data is randomly sampled to construct candidate planes, and the fitness of each candidate plane is calculated. By comparing the fitting effects of different planes through multiple iterations, the plane with the best fitting effect is selected as the ground point cloud model. This ground model can effectively distinguish itself from other equipment point clouds. Finally, using the fitted ground point cloud model, ground-related point cloud data is deleted, thereby removing ground information and generating a point cloud dataset containing only substation equipment. This dataset undergoes denoising, registration, and ground point cloud deletion steps to ensure data accuracy and validity, providing high-quality point cloud data support for subsequent equipment identification and classification.

[0041] Furthermore, the ground model fitting module includes:

[0042] The 3D point cloud dataset is preprocessed. M points are randomly sampled from the preprocessed point cloud dataset to construct a first candidate plane, where M is an integer greater than or equal to 3. The distance from each point in the preprocessed point cloud dataset to the first candidate plane is calculated, and the fitness of the first plane is calculated. A second candidate plane, different from the first candidate plane, is constructed, and its fitness is calculated. The fitness of the first plane and the fitness of the second plane are compared to determine a temporary optimal fitting plane, and a preferred direction for the first plane is constructed. A third candidate plane is constructed based on the preferred direction of the first plane, and its fitness is calculated. Iterative fitting is performed by combining the temporary optimal fitting plane and its fitness until convergence is achieved, and the candidate plane with the largest fitness is obtained to construct the ground point cloud model. The point cloud data in the ground point cloud model is deleted from the preprocessed point cloud dataset to generate the substation point cloud dataset.

[0043] In a preferred embodiment, after preprocessing the 3D point cloud dataset in the same manner as described above, the preprocessed point cloud data is randomly sampled to select M points (M≥3) to construct a first candidate plane. The purpose of random sampling is to select a subset of points from the entire point cloud to calculate and fit a plane model. Subsequently, the selected M points are used for least squares fitting to calculate the plane equation, obtain the position and orientation of the plane, and construct the first candidate plane, which represents a preliminary model of the ground point cloud. After obtaining the first candidate plane, the distance from each point in the preprocessed point cloud dataset to the plane is calculated. These distances can be used to evaluate the quality of the plane fitting. Based on these distances and a preset distance threshold, the fitness of the first candidate plane is calculated. The higher the fitness, the more representative the plane is of the ground point cloud. Next, another set of M points is randomly selected from the point cloud dataset to construct a second candidate plane, different from the first candidate plane. The fitness of the second candidate plane is then evaluated using the same calculation method and compared with the fitting effect of the first candidate plane. The plane with higher fitness is selected as the temporary optimal fitting plane. Based on the normal and translation vectors of the temporary optimal fitting plane, the preferred direction of the plane is determined to guide subsequent plane fitting and improve accuracy. Then, based on the preferred direction of the first plane, a third candidate plane is constructed, and its fitness is calculated iteratively. Each iteration combines the current optimal fitting plane and fitness to optimize the next round of candidate planes. Through multiple iterations, new candidate planes are continuously constructed and their fitness is calculated until convergence is reached (e.g., fitness no longer significantly improves or the maximum number of iterations is reached). After iterative fitting, the candidate plane with the highest fitness is selected as the final ground point cloud model. Finally, based on this ground point cloud model, point cloud data belonging to the ground is identified and deleted. These point cloud data are typically located near the fitted plane and within a preset threshold. After deleting ground point cloud data, the system terminal generates a point cloud dataset containing only substation equipment. This dataset removes ground interference while retaining the point cloud information of the substation equipment, providing a high-quality data foundation for subsequent equipment segmentation and classification. In summary, through random sampling, plane fitting, fitness calculation, and iterative optimization, ground point cloud data can be accurately identified and deleted to generate a high-quality substation point cloud dataset. This process not only improves the accuracy of the point cloud data but also lays a solid foundation for subsequent equipment identification and classification.

[0044] Furthermore, this application provides methods for calculating the distance from each point in the preprocessed point cloud dataset to the first candidate plane, and calculating the fitness of the first plane fit, including:

[0045] Calculate the distance from each point in the preprocessed point cloud dataset to the first candidate plane to obtain each distance value; based on each distance value, count the proportion of points in the preprocessed point cloud dataset whose distance value is less than a preset distance threshold, and generate the first plane fitting fitness.

[0046] In one optional implementation, when calculating the distance from each point in the preprocessed point cloud dataset to the first candidate plane, the distance from each point in the point cloud to the plane is first calculated based on the fitted equation of the first candidate plane. These distances can be calculated using the point-to-plane distance formula, specifically:

[0047] Among them, (x i ,y i ,z i Let be the coordinates of the i-th point in the point cloud; A, B, C, and D are the equation parameters of the first candidate plane, representing the coefficients in the plane equation Ax + By + Cz + D = 0. Using the above formula, the distance d from each point to the first candidate plane can be calculated. i Subsequently, based on these distance values, the number of points in the point cloud dataset whose distance values ​​are less than a preset distance threshold is counted. This preset distance threshold is a very small value, generally understood as a ground threshold used to determine whether a point belongs to the ground point cloud. The fitness of the first candidate plane is calculated by dividing the number of points with distances less than the preset threshold by the total number of points in the point cloud data. This fitness reflects the degree to which the plane fits the ground point cloud data; a higher fitness indicates that the plane accurately represents the ground, further confirming its effectiveness as a ground point cloud model. Finally, based on this fitness value, the effectiveness of the candidate plane can be evaluated, deciding whether to further optimize or replace it.

[0048] For example, in a preprocessed point cloud dataset with 1000 points and a preset distance threshold of 0.1 meters, after calculating the distance of each point to the first candidate plane, it is found that the distance of 800 points is less than 0.1 meters, so the fitness of the first plane is 0.8.

[0049] Furthermore, this application provides a method for comparing the fitness of the first plane fitting with the fitness of the second plane fitting, determining a temporary optimal fitting plane, and constructing a preferred direction for the first plane, including:

[0050] The fitness of the first plane fitting is compared with that of the second plane fitting. The candidate plane with the higher fitness is selected as the temporary optimal fitting plane, and the candidate plane with the lower fitness is selected as the taboo fitting plane. The trend of the change of the normal vector and the translation vector of the taboo fitting plane to the temporary optimal fitting plane are analyzed to generate the preferred direction of the first plane.

[0051] In one optional implementation, the system terminal calculates the fitness of the first and second planes respectively. The fitness is quantified by evaluating the matching degree between the plane and the point cloud data. If the fitness of the first plane is greater than that of the second plane, the first plane has a better fit and becomes the temporary optimal fitting plane, while the second candidate plane is designated as a taboo fitting plane, meaning it is not selected in the current iteration. If the fitness of the first plane is less than that of the second plane, the second plane has a better fit and becomes the temporary optimal fitting plane, while the first candidate plane is designated as a taboo fitting plane. Subsequently, the system terminal analyzes the trend of the normal vector changes of the taboo fitting plane and the temporary optimal fitting plane. The normal vector represents the direction of the plane, and the trend reflects the adjustment of the plane's direction. The normal vector is a coefficient in the plane equation, representing the plane's orientation. For example, the normal vector of the tabu-fitting plane is n1 = (A1, B1, C1), and the normal vector of the temporary optimal-fitting plane is n2 = (A2, B2, C2). The trend of normal vector change can be represented by the vector difference Δn = n2 - n1. The translation vector represents the adjustment of the plane's position. For example, the translation vector of the tabu-fitting plane is t1 = (D1), and the normal vector of the temporary optimal-fitting plane is t2 = (D2). The change of translation vector can be represented by the vector difference Δt = t2 - t1. Finally, based on the analysis of the trends of normal vector change and translation vectors of the tabu-fitting plane and the temporary optimal-fitting plane, a preferred direction for the first plane can be generated. This preferred direction will guide the subsequent plane fitting process, improving the efficiency of plane fitting and helping to find a better fitting plane more quickly.

[0052] The density clustering analysis module 13 is used to perform density clustering segmentation on the substation cloud dataset to generate the first point cloud segmentation result.

[0053] Specifically, in the density clustering analysis module 13, when performing density clustering segmentation on the substation point cloud dataset, the distribution density of the point cloud data is first analyzed. The purpose of density clustering is to group points that are close to each other in the point cloud into a cluster. This is typically determined by setting a neighborhood radius and a minimum number of neighboring points. Specifically, a threshold for the neighborhood radius and the minimum number of neighboring points is set. The neighborhood radius determines the maximum distance a point can reach in space, while the minimum number of neighboring points refers to the minimum number of points required to form a cluster within the neighborhood radius. Then, the density clustering algorithm starts with a point in the point cloud data and searches for all its neighboring points within the neighborhood radius. If the number of points in the neighborhood is greater than or equal to the minimum number of neighboring points, these points are grouped into the same cluster. The algorithm then processes other points in the neighborhood of these points in a similar way until the entire point cloud data is segmented into several clusters. In this way, the point cloud dataset is segmented into multiple clusters with similar spatial distributions, each cluster representing an independent device or structural component. This process is called first-point cloud segmentation, and the resulting first-point cloud segmentation results contain the spatial division between various devices in the substation. Ultimately, these segmented clusters can be further used for device identification, classification, and analysis. Density clustering segmentation not only effectively identifies the boundaries between different devices but also avoids the problems of missegmentation or omissions that may occur in traditional methods, especially when there are complex shapes and relatively close spatial arrangements between devices.

[0054] Furthermore, the density clustering analysis module includes:

[0055] Analyze the point cloud distribution density in the substation site cloud dataset, configure the neighborhood radius and minimum number of neighborhood points; perform density clustering based on the neighborhood radius and minimum number of neighborhood points, starting from any core point, group all density-reachable points into the same cluster, complete the density clustering segmentation, and generate the first point cloud segmentation result.

[0056] In a preferred embodiment, when performing density clustering segmentation of a substation point cloud dataset, it is first necessary to analyze the point cloud distribution density. To achieve this, the system terminal performs density clustering by configuring the neighborhood radius (∈) and the minimum number of neighborhood points (MinPts). The minimum number of neighborhood points is a key parameter in the density clustering algorithm, used to determine the minimum number of points in the neighborhood of a given point to be considered a core point. In 3D point cloud data processing, MinPts is typically between 4 and 10, and the specific value can be adjusted according to the density of the point cloud. Common choices are 4 or 5. Smaller values ​​are suitable for denser point clouds, while larger values ​​are suitable for sparser point clouds. The neighborhood radius determines how the neighborhood range is defined in the clustering algorithm; that is, the maximum distance around a point that can be considered its neighbors. To choose a suitable ∈ value, the system terminal calculates the distance from each point in the dataset to its k-th nearest neighbor, where k = MinPts - 1. For example, if MinPts = 5, the distance from each point to its 4th nearest neighbor is calculated. Based on these distance values, a k-distance graph (distance and ranking graph) is plotted. This graph shows the distance distribution between a point and its k-th nearest neighbor. By observing the k-distance graph, inflection points are found in the graph. The location of an inflection point usually corresponds to a suitable ∈ value, which represents the range of dense regions in the point cloud. Points with a distance near this value should be grouped into the same cluster. In addition, if the average spacing of the point cloud is known, ∈ can be set using a multiple of this value, usually set to 1 to 2 times the average spacing of the point cloud. After determining ∈ and MinPts, the density clustering algorithm starts clustering from any core point in the point cloud data. A core point is defined as one whose neighborhood contains at least MinPts points. For each core point, all points within its neighborhood are traversed, and density-reachable points are grouped into the same cluster. If a core point exists within the neighborhood of a point, and the distance to that point is less than ∈ , then that point is considered density-reachable and belongs to the cluster of the current core point. In this way, the density clustering algorithm continuously expands the clusters until all core points and their neighborhood points are assigned to a cluster; points that cannot be expanded are marked as noise. Ultimately, this process completes the segmentation of the point cloud data and generates the first point cloud segmentation result. Through density clustering segmentation, the substation's point cloud data is divided into multiple clusters, each representing an independent equipment or object region. The first point cloud segmentation result is the set of these independent clusters, and each cluster can be further used for equipment identification, classification, or other analysis tasks. This process ensures that high-density point cloud regions can be correctly classified as the same equipment or object and effectively avoids noise interference in the point cloud.

[0057] The segmentation boundary optimization module is used to perform normal vector and curvature similarity analysis on the segmentation boundary of the first point cloud segmentation result, optimize the segmentation boundary, and generate point cloud individual segmentation results.

[0058] Specifically, in the segmentation boundary optimization module 14, after the first point cloud segmentation, the generated point cloud clusters may have some boundary regions. These boundaries may be caused by the irregularity of the device shape or errors in the point cloud acquisition process. Therefore, further optimization is needed for these segmentation boundaries to improve the accuracy and reliability of the segmentation results. To optimize the segmentation boundaries, the normal vector and curvature information of the boundary points must first be analyzed. The normal vector is a vector representing the surface orientation. In point cloud processing, the normal vector is usually used to describe the surface orientation of each point in the point cloud. Curvature is a parameter describing the degree of curvature of the point cloud surface and is usually used to identify the geometric features of the point cloud surface. For adjacent boundary points, the angle between their normal vectors is calculated. If the normal vector directions of adjacent points are similar, it means that these points are likely to belong to the surface of the same device or object and can be merged into the same cluster. Conversely, if the normal vector directions are significantly different, it means that these points may belong to the boundaries of different devices or objects and need to be segmented. Similarly, curvature can be used to evaluate the shape similarity between adjacent points. If two adjacent points have similar curvature values, it indicates they are located in similar surface regions and belong to the boundary of the same object, and can be merged into a cluster. If the curvature differences are large, it means the two points may belong to different surfaces and should be segmented. For adjacent boundary points with similar normal vectors and curvature, grouping them into the same cluster avoids missegmentation caused by errors in point cloud acquisition or irregularities in object shape. For adjacent boundary points with large differences in normal vectors and curvature, separating them from the original clusters generates new independent clusters. This ensures that each cluster represents an independent object or device part. Boundary optimization after normal vector and curvature similarity analysis generates more accurate point cloud unit segmentation results. The optimized point cloud unit segmentation results ensure that each independent device or object is correctly separated, eliminating missegmentation and missed segmentation problems. These unit segmentation results can serve as the basis for subsequent device identification and classification. In summary, normal vector and curvature similarity analysis of segmentation boundaries, as well as the boundary optimization process, can significantly improve the accuracy of point cloud data segmentation, ensuring the independence and integrity of each device or object in the point cloud data.

[0059] Furthermore, the density clustering analysis module includes:

[0060] Extract the first and second preliminary segmentation clusters that are adjacent in position from the first point cloud segmentation result, as well as the corresponding first and second adjacent boundary point cloud data. Perform stability analysis on the normal vector and curvature information of each point in the first and second adjacent boundary point cloud data respectively to generate a first boundary stability evaluation value and a second boundary stability evaluation value. Determine whether the first boundary stability evaluation value and the second boundary stability evaluation value both meet the preset boundary stability evaluation threshold. If not, re-segment the first and second adjacent boundary point cloud data until the re-segmented first and second boundary stability evaluation values ​​both meet the preset boundary stability evaluation threshold. Generate the point cloud individual segmentation result based on the re-segmentation result.

[0061] In a preferred embodiment, when optimizing the first point cloud segmentation result, firstly, adjacent first and second preliminary segmentation clusters are extracted from the first point cloud segmentation result. These clusters represent parts of the point cloud data that may belong to different devices or objects. For each pair of adjacent clusters, their adjacent boundary point cloud data, i.e., the boundary parts of the two clusters, are also extracted. These boundary points are usually located in the transition region between clusters. Subsequently, the system terminal calculates the normal vector of each point in the first and second adjacent boundary point cloud data, and calculates the angle between the normal vectors of adjacent points. The smaller the angle between the normal vectors, the more consistent the surface orientation of the points, and the higher the stability. At the same time, the curvature value of each point is calculated. The curvature value reflects the degree of curvature of the point cloud surface. The smaller the curvature value, the flatter the surface of the point, and the higher the stability. The formula for calculating curvature k is: Where λ0, λ1, and λ2 are the eigenvalues ​​of the point neighborhood covariance matrix, and λ0 ≤ λ1 ≤ λ2. Then, for the first adjacent boundary point cloud data, the stability of its normal vector angle and curvature values ​​is calculated using a weighted method to generate a first boundary stability evaluation value. Similarly, for the second adjacent boundary point cloud data, the stability of its normal vector and curvature is calculated to generate a second boundary stability evaluation value. These stability evaluation values ​​are used to quantify the smoothness and consistency of each boundary; a higher stability evaluation value indicates a more stable boundary, while a lower value suggests a potential problem. Once the stability evaluation values ​​of the first and second boundaries are obtained, they need to be compared with a preset boundary stability evaluation threshold. This threshold is usually determined experimentally or empirically to determine whether the boundary is sufficiently stable. If both the first and second boundary stability evaluation values ​​meet the preset threshold, it means that the two boundaries are stable and the current segmentation result can be maintained. If the stability evaluation value of either boundary does not reach the preset threshold, the two adjacent boundary point cloud data need to be re-segmented to ensure that the generated boundaries are stable and accurate. When resegmentation is triggered, the system terminal adjusts the boundary positions based on information such as normal vectors and curvature to eliminate the effects of missegmentation or noise. This process is repeated multiple times until both the first and second boundary stability evaluation values ​​meet the preset stability threshold. Finally, the optimized segmentation results are integrated to generate individual point cloud segmentation results. These results accurately represent the equipment or object regions within the substation, ensuring that the point cloud data of each device is accurately separated and identified, thus providing reliable point cloud data for subsequent equipment identification and classification.

[0062] Furthermore, this application provides a method for re-segmenting the boundary of the first adjacent boundary point cloud data and the second adjacent boundary point cloud data, including:

[0063] Locate the locations of abnormal normal vector and curvature changes in the first and second adjacent boundary point cloud data; re-segment the boundaries based on the locations of abnormal normal vector and curvature changes, and iterate and optimize in this way until the stability evaluation values ​​of the first and second boundaries after re-segmentation both meet the preset boundary stability evaluation threshold, and generate the point cloud unit segmentation result based on the re-segmentation result.

[0064] In one optional implementation, when optimizing the boundary of point cloud data, it is first necessary to locate the abnormal positions of normal vector and curvature changes in the first and second adjacent boundary point cloud data. Specifically, the normal vector and curvature values ​​of each boundary point are analyzed to locate the abnormal positions. Abnormal positions are usually characterized by drastic changes in normal vector or large fluctuations in curvature value, i.e., the angle between the normal vector and the neighboring point is greater than the corresponding threshold or the difference in curvature value is greater than the corresponding threshold. These abnormal changes usually occur in the contact or segmentation areas of different objects, which may be a manifestation of missegmentation or boundary instability. Based on these analyses, the system terminal will locate the boundary points with abnormal changes in normal vector and curvature and record their position information for subsequent resegmentation. Once these abnormal points are located, the system terminal will use Euclidean clustering algorithm to reseg the abnormal points. The core of the Euclidean clustering algorithm is to build a kd-tree to accelerate nearest neighbor search and ensure that each device is correctly separated. A kd-tree is a spatial partitioning data structure that can efficiently find the nearest neighbor points in point cloud data. The system terminal starts from any core point, groups all points within its neighborhood into the same cluster, and then checks if these neighboring points are also core points. If so, it adds the points within their neighborhood to the current cluster. This process is repeated recursively until no new points can be added to the current cluster. Based on the changing trends of normal vectors and curvature, the system terminal reassigns outliers to appropriate clusters. For example, if the normal vector and curvature values ​​of an outlier are more similar to the first preliminary segmentation cluster, it is assigned to the first preliminary segmentation cluster; if they are more similar to the second preliminary segmentation cluster, it is assigned to the second preliminary segmentation cluster. The system terminal repeats the above process of locating outliers and resegmenting until the first and second boundary stability evaluation values ​​after resegmentation both meet the preset boundary stability evaluation threshold. After each iteration, the normal vector and curvature stability of the first and second adjacent boundary point cloud data are recalculated to generate new boundary stability evaluation values. If the evaluation value meets the preset threshold, the iteration stops; otherwise, optimization continues. Through iterative optimization, optimized point cloud individual segmentation results are generated, and each device or object is accurately segmented into independent point cloud clusters with clearer segmentation boundaries. During the optimization process, points that cannot be assigned to any cluster are marked as noise points. These points are typically located at the edges of devices or in the background and do not belong to any device point cloud. In summary, through normal vector and curvature analysis, boundary re-segmentation, optimization of the Euclidean clustering algorithm, and iterative optimization, each device or object in the point cloud data can be accurately segmented, eliminating missegmentation and ensuring the accuracy of the final individual point cloud segmentation results. This process improves segmentation accuracy and provides a reliable foundation for subsequent device identification and classification.

[0065] The device classification module 15 is used to perform multi-view 2D image rendering on the point cloud single-unit segmentation results, and to classify devices based on the 2D image rendering results.

[0066] Specifically, in the device classification module 15, after completing the point cloud individual segmentation, multi-view 2D image rendering is required for the point cloud data of each device or object. The goal of multi-view rendering is to display the 3D point cloud data of each device or object as a 2D image from different perspectives. During this process, the optimal perspective for observing each device or object is selected by calculating the information entropy of each perspective. For the selected perspective, the 3D point cloud of the device is projected onto a 2D plane, and point cloud rendering technology is used to convert the spatial information of the point cloud data into image information, generating the corresponding 2D image. Each perspective image shows a different view of the device, which helps in subsequent analysis and recognition. After rendering, the multi-view 2D images are used as input for subsequent deep learning models. The deep learning model extracts key visual features from the rendered 2D images, such as edges, shapes, and textures, and identifies the type of device based on existing training data. For different types of devices, the model classifies them based on these visual features. Finally, the classification results output a category label for each device, indicating whether the device belongs to a transformer, switchgear, circuit breaker, or other similar device type. In summary, multi-view 2D image rendering not only transforms 3D point cloud data into 2D images, allowing different perspectives of the equipment to be displayed, but also provides rich visual information for subsequent equipment classification. Combined with machine learning models, it can effectively identify and classify various types of equipment in substations, improving the automation and accuracy of equipment management.

[0067] When constructing a deep learning model for substation equipment classification, ResNet-18 or EfficientNet are first used as the base model. A deep neural network structure is built, consisting of an input layer, multiple convolutional layers, residual blocks, a global average pooling layer, and a fully connected layer. The first layer is a 7×7 convolutional layer with 64 channels and a stride of 2 to extract preliminary image features. Subsequently, a 3×3 max pooling layer downsamples the feature map to reduce computation and enhance feature representation. The network extracts deep features through four residual blocks, each containing two convolutional layers. Residual connections are used to avoid the vanishing gradient problem, enabling the network to effectively learn deep features of equipment categories. After feature extraction through the residual blocks, the network enters a global average pooling layer to further compress the feature map size, and finally, the fully connected layer outputs the equipment classification results. Before model training, image datasets of substation equipment were collected and labeled. These images were generated through multi-view 2D rendering of point cloud data, covering different angles and states of the equipment. To improve model training efficiency, transfer learning techniques were employed to fine-tune the existing pre-trained model. Specifically, pre-trained ResNet-18 or EfficientNet weights were loaded, and the parameters of the first few layers were frozen to maintain their original feature extraction capabilities. The last fully connected layer of the model was then replaced with a new fully connected layer to match the number of equipment categories. Afterward, the network was fine-tuned using a small learning rate to ensure that the model could learn the specific features of substation equipment while retaining general features. During training, the exported multi-view 2D images were batch-inputted. The model sequentially passed through the input layer, convolutional layer, residual block, global average pooling layer, and fully connected layer for forward propagation, calculating the predicted probability for each equipment category. The cross-entropy loss function was then used to calculate the loss between the predicted result and the true label, and backpropagation was used to calculate the gradient of the loss with respect to the weights of each layer. Finally, the Adam optimizer was used to optimize the model parameters, adjusting the network weights to minimize the loss value. This training process is repeated until the maximum number of iterations is reached or the model's classification accuracy converges to the expected range. After training, validation data is used to evaluate the model's performance and test its classification accuracy for different device categories. If the model's accuracy on the device classification task meets expectations, the trained model is used as the final classifier for device recognition. Otherwise, hyperparameters such as the learning rate and the number of training batches are adjusted to further optimize the model's performance.

[0068] Furthermore, the equipment classification module includes:

[0069] Each device unit in the point cloud unit segmentation result is rotated 360 degrees, and the information entropy of each viewpoint is calculated. Based on the information entropy of each viewpoint, the best viewpoint is automatically selected according to the principle of maximizing information entropy, and 2D image rendering is performed to export the 2D image.

[0070] In a preferred embodiment, before rendering 2D images of the point cloud unit segmentation results, each unit needs to be rotated 360 degrees to generate multiple viewpoint projection images of the point cloud from different angles. The goal of this process is to find the optimal viewpoint that best showcases the structural features of the device, thereby improving the accuracy of subsequent classification. During the rotation of the unit, samples are taken at multiple preset angles, and a 2D projection image is generated at each viewpoint. To quantify the information richness of each viewpoint, entropy is used as a metric. The formula for calculating entropy is as follows: Where H represents the information entropy from the current perspective; p i is the normalized intensity value of the i-th pixel in the image, obtained by normalizing the pixel's grayscale value according to its maximum and minimum values; n is the total number of pixels in the image. After calculating the information entropy of each viewpoint, the system terminal automatically selects the optimal viewpoint based on the principle of maximizing information entropy. The core idea of ​​this principle is to select the viewpoint with the highest information content as the final 2D image rendering angle, as these angles provide the richest feature information. After determining the optimal viewpoint, the system terminal performs 2D projection from that viewpoint, converting the point cloud data into a 2D image. During the rendering process, methods such as depth mapping and shadow processing may be used to enhance the image's visualization effect, making it suitable for subsequent classification tasks. Finally, the exported 2D image will be used for automatic equipment classification and further optimize the identification and management process of substation equipment. This information entropy-driven viewpoint selection method ensures that the rendered image contains the richest feature information, providing high-quality data support for the inference of deep learning models.

[0071] In summary, the intelligent classification system for outdoor substation equipment based on point cloud data recognition provided in this application has the following technical advantages:

[0072] This application uses a 3D laser scanner to perform a 3D all-around scan of the target outdoor substation using an echo analysis module 11, and performs multi-echo analysis of laser pulses to establish a 3D point cloud dataset after equipment overlap and segmentation. The ground model fitting module 12 preprocesses the 3D point cloud dataset, then iteratively fits the ground point cloud model, and deletes ground point cloud data based on the fitting results to generate a substation site cloud dataset. The density clustering analysis module 13 performs density clustering segmentation on the substation site cloud dataset to generate a first point cloud segmentation result. The segmentation boundary optimization module 14 performs normal vector and curvature similarity analysis on the segmentation boundary of the first point cloud segmentation result to optimize the segmentation boundary and generate individual point cloud segmentation results. The equipment classification module 15 performs multi-view 2D image rendering on the individual point cloud segmentation results, and classifies the equipment based on the 2D image rendering results. These technologies collectively solve the technical problems in traditional point cloud processing, such as low equipment recognition efficiency due to reliance on manual segmentation, accumulation of multi-source point cloud registration errors, and misclassification caused by blurred segmentation boundaries. They achieve the technical effects of fully automatic high-precision classification of substation equipment by optimizing point cloud acquisition accuracy based on multi-echo analysis, achieving high robust segmentation through iterative fitting and density clustering, and improving boundary definition accuracy by fusing normal vectors and curvature analysis.

[0073] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

[0074] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of this application and its equivalents, this application also intends to include such modifications and variations.

Claims

1. An intelligent classification system for outdoor substation equipment based on point cloud data recognition, characterized in that: include: The echo analysis module is used to perform omnidirectional scanning of the target outdoor substation using a 3D laser scanner and to perform multi-echo analysis of laser pulses to establish a 3D point cloud dataset after equipment overlap segmentation. The ground model fitting module is used to preprocess the 3D point cloud dataset, perform iterative fitting of the ground point cloud model, delete ground point cloud data based on the fitting results, and generate a substation site cloud dataset. The density clustering analysis module is used to perform density clustering segmentation on the substation site cloud dataset to generate the first point cloud segmentation result. The segmentation boundary optimization module is used to perform normal vector and curvature similarity analysis on the segmentation boundary of the first point cloud segmentation result, optimize the segmentation boundary, and generate point cloud unit segmentation results; the device classification module is used to perform multi-view 2D image rendering on the point cloud unit segmentation results, and classify devices based on the 2D image rendering results; wherein, the ground model fitting module includes: preprocessing the three-dimensional point cloud dataset, randomly sampling M points in the preprocessed point cloud dataset to construct a first candidate plane, where M is an integer greater than or equal to 3; calculating the distance from each point in the preprocessed point cloud dataset to the first candidate plane, and calculating the fitness of the first plane fitting; Continue constructing a second candidate plane different from the first candidate plane and calculate the fitness of the second plane fitting; compare the fitness of the first plane fitting with the fitness of the second plane fitting to determine the temporary optimal fitting plane and construct the preferred direction of the first plane; construct a third candidate plane based on the preferred direction of the first plane and calculate the fitness of the third plane fitting, and perform iterative fitting by combining the temporary optimal fitting plane and the fitness of the temporary optimal plane fitting until the convergence condition is reached, obtain the candidate plane with the largest plane fitting fitness, and construct the ground point cloud model; delete the point cloud data in the ground point cloud model from the preprocessed point cloud dataset to generate the substation point cloud dataset. The calculation of the distance from each point in the preprocessed point cloud dataset to the first candidate plane, and the calculation of the first plane fitting fitness, includes: calculating the distance from each point in the preprocessed point cloud dataset to the first candidate plane, obtaining each distance value; based on each distance value, statistically analyzing the proportion of points in the preprocessed point cloud dataset whose distance values ​​are less than a preset distance threshold, and generating the first plane fitting fitness; and comparing the first plane fitting fitness with the second plane fitting fitness to determine a temporary optimal fitting plane and constructing a first plane preferred direction, including: comparing the first plane fitting fitness with the second plane fitting fitness, and selecting the candidate plane with the higher fitness. The plane is used as the temporary optimal fitting plane, and the candidate plane with low fitness is used as the taboo fitting plane; the change trend of the normal vector and the translation vector of the taboo fitting plane to the temporary optimal fitting plane are analyzed to generate the preferred direction of the first plane; wherein, the density clustering analysis module includes: extracting the first preliminary segmentation cluster and the second preliminary segmentation cluster that are adjacent in position from the first point cloud segmentation result, as well as the corresponding first adjacent boundary point cloud data and the second adjacent boundary point cloud data; performing stability analysis on the normal vector and curvature information of each point in the first adjacent boundary point cloud data and the second adjacent boundary point cloud data respectively, and generating the first boundary stability evaluation value and the second boundary stability evaluation value.The system determines whether both the first boundary stability evaluation value and the second boundary stability evaluation value meet a preset boundary stability evaluation threshold. If not, it re-segments the first and second adjacent boundary point cloud data until both the re-segmented first and second boundary stability evaluation values ​​meet the preset boundary stability evaluation threshold. The re-segmentation result is then used to generate the point cloud unit segmentation result. The re-segmentation of the first and second adjacent boundary point cloud data includes: locating locations in the first and second adjacent boundary point cloud data where normal vectors and curvature changes are abnormal; re-segmenting the boundary based on these abnormal locations; and iteratively optimizing in this manner until both the re-segmented first and second boundary stability evaluation values ​​meet the preset boundary stability evaluation threshold. The re-segmentation result is then used to generate the point cloud unit segmentation result.

2. The intelligent classification system for outdoor substation equipment combining point cloud data recognition as described in claim 1, characterized in that, The echo analysis module further includes: receiving laser pulse echo data from each scan performed by the 3D laser scanner; extracting multiple echo data from the laser pulse echo data; calculating the echo spacing between adjacent first and second echo data in the multiple echo data to generate a first echo spacing; determining whether the first echo spacing is greater than a preset spacing threshold; if so, performing preliminary point cloud overlap segmentation based on the first echo data and the second echo data to generate segmented point cloud data, and adding it to the 3D point cloud dataset.

3. The intelligent classification system for outdoor substation equipment combining point cloud data recognition as described in claim 2, characterized in that, Determining whether the first echo spacing is greater than a preset spacing threshold includes: if the first echo spacing is less than or equal to the preset spacing threshold, merging the first echo data and the second echo data, and adding the processed point cloud into the three-dimensional point cloud dataset.

4. The intelligent classification system for outdoor substation equipment combining point cloud data recognition as described in claim 1, characterized in that, The density clustering analysis module includes: analyzing the point cloud distribution density in the substation site cloud dataset, configuring the neighborhood radius and the minimum number of neighborhood points; performing density clustering based on the neighborhood radius and the minimum number of neighborhood points, starting from any core point, grouping all density-reachable points into the same cluster, completing the density clustering segmentation, and generating the first point cloud segmentation result.

5. The intelligent classification system for outdoor substation equipment combining point cloud data recognition as described in claim 1, characterized in that, The device classification module includes: rotating each device in the point cloud individual segmentation result by 360 degrees, calculating the information entropy of each viewpoint; based on the information entropy of each viewpoint, automatically selecting the best viewpoint according to the principle of maximizing information entropy, performing 2D image rendering, and exporting the 2D image.

Images