GIS equipment deformation detection method based on point cloud registration

Abstract

According to the GIS equipment deformation detection method based on point cloud registration, a lightweight hardware architecture is adopted, a high-precision three-dimensional imaging system and an embedded processing unit are integrated, and point cloud data acquisition, multi-stage registration, deformation analysis and risk early warning are realized in combination with a software system. According to the method, a curvature constraint-based multi-stage point cloud registration algorithm is provided, and the registration precision is improved through adjustable farthest point sampling, maximum clique preliminary registration and curvature feature constraint fine registration; an innovative improved point cloud comparison method is combined with octree search and a local Hausdorff distance to realize accurate deformation detection. The problems that a traditional monitoring method is low in precision, poor in efficiency and insufficient in portability are solved, automatic and high-precision deformation monitoring of GIS equipment can be achieved, and guarantee is provided for safe operation of a power grid.

Description

Technical Field

[0001] This invention relates to the technical field of cross-application of power equipment condition monitoring and intelligent operation and maintenance, specifically to a GIS equipment deformation detection method based on point cloud registration. Background Technology

[0002] Gas-insulated switchgear (GIS) has become a core power equipment in modern substations due to its compact structure, small footprint, and high reliability. However, during long-term operation, GIS equipment is susceptible to problems such as tank deformation and seal failure due to temperature gradients, mechanical stress, environmental corrosion, and operating voltage. In severe cases, this can even lead to electrical breakdown and equipment failure, posing a risk to the safe operation of the power grid. Currently, commonly used methods for detecting deformation in GIS equipment mainly include: Manual inspection method: This method involves manually inspecting equipment for deformation using measuring tools or by visual inspection. While low-cost, it suffers from drawbacks such as high labor intensity, low accuracy, inability to perform continuous monitoring, and the safety risks associated with high-pressure operations on-site.

[0003] Ultrasonic testing: This method monitors localized deformation of equipment using ultrasonic sensors. While it can measure localized structures with high precision, its detection range is limited, requiring a large number of sensors to achieve full coverage, leading to a significant increase in integration complexity and cost.

[0004] Traditional machine vision and point cloud analysis methods utilize 2D images or conventional point cloud registration methods for deformation detection. Existing methods generally suffer from the following problems: sensitivity to initial pose, susceptibility to local optima, and insufficient registration accuracy; the easy loss of key feature points when processing complex curved surface point cloud data, affecting overall registration quality; lack of robustness to point cloud noise and local distortion, leading to unstable deformation region detection; and difficulty in incorporating environmental factors (such as temperature) for comprehensive risk assessment.

[0005] Patent application CN110256347A discloses a laser scanning-based deformation detection method for power equipment. This method uses a laser scanner to collect point cloud data of the equipment surface and then calculates the deformation using a point-to-point fitting algorithm. While this approach is effective for small-scale, high-precision detection, the high cost of laser scanning equipment and its sensitivity to external lighting and surface reflectivity make it difficult to apply stably in the complex environment of outdoor power grids over long periods.

[0006] Patent application CN113785902A discloses a GIS equipment condition detection system based on infrared thermal imaging. This system uses a thermal imager to monitor the surface temperature distribution of the equipment in real time and combines image analysis methods to determine whether there are any abnormalities in the equipment. This method can quickly detect potential defects caused by heat, but it can only reflect surface temperature changes and cannot directly quantify the deformation of the equipment's geometric structure, resulting in insufficient detection accuracy.

[0007] Patent application CN114659021A discloses a deformation monitoring device for power equipment that combines an accelerometer and a strain gauge. It collects dynamic vibration signals from the equipment through a multi-point sensor deployment and estimates deformation through feature extraction and stress-strain calculation. While this method has some detection capability for deformation caused by dynamic impact or vibration, it requires deploying a large number of sensors on the equipment surface, compromising the sealing and integrity of the GIS equipment, and also involves complex wiring and difficult maintenance. Summary of the Invention

[0008] The technical problem to be solved by this invention is: how to achieve accurate detection of deformation of GIS equipment through point cloud registration when sample data is scarce.

[0009] The present invention proposes a point cloud registration-based deformation detection method for GIS equipment, which achieves high-precision three-dimensional deformation monitoring and risk prediction of gas-insulated switchgear (GIS) through multi-stage point cloud registration and an improved point cloud comparison algorithm.

[0010] To achieve the above objectives, the present invention adopts the following technical solution: A method for deformation detection of GIS equipment based on point cloud registration includes the following steps: S1: Data Acquisition Use a binocular camera to acquire 3D point cloud data from GIS devices; S2: Data Preprocessing The acquired 3D point cloud data is preprocessed to obtain preprocessed point cloud data; S3: Point Cloud Registration Preliminary registration of point cloud data is performed using feature descriptors PN combined with the RANSAC algorithm, and then precise registration of point cloud data is performed using the ICP algorithm accelerated by the KD tree nearest neighbor search method. S4: Deformation Detection Based on the point cloud registration results, the displacement distance between the measured point cloud and the ideal point cloud is calculated, and the deformation threshold is determined using the deformation threshold formula. By changing the deformation threshold By comparing with displacement distance, all deformation points are identified and integrated to determine the deformation area of ​​the GIS equipment; S5: Model Predictive Analysis The detected deformed areas are input into a trained prediction model, which then performs detailed analysis and risk assessment of the deformed areas.

[0011] Furthermore, in step S2, the data preprocessing process is as follows: S21: Farthest point sampling The farthest point sampling operation is performed on the 3D point cloud data. A point is randomly selected from the point cloud as the initial point and added to the sampling point set. Then, the farthest point is selected and added to the sampling point set until the required number of sampling points is reached and sampling stops. S22: Filtering Process Calculate the average distance and standard deviation between each sampling point in the point cloud and its neighboring points, and remove points whose distance exceeds a set threshold to remove isolated and noisy points from the data; for each point in the point cloud, calculate its weighted average with its neighboring points, with the weights determined by a Gaussian filter, to update the value of that point.

[0012] Furthermore, in step S3, the specific process of preliminary registration is as follows: S301: Use the spherical distance metric to find the neighborhood points corresponding to each sampling point in the preprocessed point cloud data; S302: Perform multilayer perceptron operations on each center point and its neighboring points to convert each point cloud data into a 128-dimensional feature vector, which is the feature descriptor PN. S303: Then, the feature descriptor PN is used for RANSA algorithm matching to establish an approximate correspondence, thereby achieving the initial alignment between the preprocessed measured point cloud and the ideal point cloud.

[0013] Furthermore, in step S3, the RANSAC algorithm can extract valid data points that conform to the model from a dataset containing multiple outliers, thereby obtaining reliable model parameter estimates. The specific process is as follows: S3011: Randomly select a subset of multiple point pairs and calculate the initial transformation matrix; S3012: Use the transformation matrix to perform a preliminary transformation on the point cloud, and evaluate it against the ideal point cloud to observe whether it meets the requirement for the number of points; S3011: Repeat the above steps until consistency is maximized or the expected number of iterations is reached, and output the best model.

[0014] Furthermore, in step S3, the ICP algorithm iteratively finds the nearest neighbor pair between two point clouds and calculates the optimal rigid transformation so that one point cloud can be aligned with the other point cloud.

[0015] Furthermore, in step S3, before performing precise registration of the point cloud, the KD-tree nearest neighbor search method is used to spatially divide the point cloud to be registered and the ideal point cloud to construct a kd-tree structure. The KD-tree nearest neighbor search method can reduce the time complexity of searching for the nearest point in the ICP algorithm from linear to logarithmic level, thereby improving the computation speed of the ICP algorithm.

[0016] Furthermore, in step S4, the deformation threshold The calculation formula is as follows:

[0017] in, The minimum deformation threshold for determining whether deformation has occurred; The overlap rate after precise registration; A represents the point cloud data quality, determined by real-time temperature. 1. 2 represents the ratio factor between the two factors; Other influencing factors were identified through analysis of the stability of the shooting equipment.

[0018] In another aspect, the present invention also discloses a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the method described above.

[0019] In another aspect, the present invention also discloses a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the method described above.

[0020] As can be seen from the above technical solution, the GIS equipment deformation detection method and system based on point cloud registration of the present invention, through the setting of a device, adopts a lightweight hardware architecture, integrates a high-precision three-dimensional imaging system and an embedded processing unit, and combines with a software system to realize point cloud data acquisition, multi-stage registration, deformation analysis, and risk warning. The method proposes a multi-stage point cloud registration algorithm based on curvature constraints, which improves registration accuracy through adjustable farthest point sampling, initial registration of maximal cliques, and fine registration constrained by curvature features; it also innovates and improves the point cloud comparison method, combining octree search and local Hausdorff distance to achieve accurate deformation detection. This invention solves the problems of low accuracy, poor efficiency, and insufficient portability of traditional monitoring methods, and can realize automated, high-precision deformation monitoring of GIS equipment, providing a guarantee for the safe operation of the power grid.

[0021] The present invention has the following advantages over the prior art: This invention addresses the limitations of existing detection methods, such as the inability to perform long-term detection and the limited detection range. It enables real-time imaging of deformation and displacement of GIS equipment over a wide range and generates a three-dimensional point cloud structure. Furthermore, it requires no manual intervention throughout the entire process, achieving automated deformation detection over a long period of time.

[0022] This invention addresses the problem of limited deformation sample data by using point cloud registration technology to detect deformation in GIS equipment. It registers and compares the ideal model with the measured model to detect the deformed parts of the equipment.

[0023] This invention addresses the problems of large computational load, uneven distribution of point cloud data, and presence of noisy points in registration algorithms. It employs a method that combines farthest point sampling and Gaussian filtering to process point cloud data, effectively reducing the amount of point cloud data while better covering the entire point cloud structure.

[0024] This invention addresses the problems of numerous mismatches and low registration efficiency in traditional preliminary registration methods, as well as the high requirements of the ICP algorithm for the initial position of the registration point cloud. It proposes a preliminary registration algorithm based on a novel feature descriptor PN combined with the iterative nearest point ICP algorithm.

[0025] This invention can perform deformation detection and analysis on the collected point cloud data and predict whether there are potential risks in the subsequent operation of GIS equipment.

[0026] This invention addresses the problems of mainstream GIS shell defect detection systems, which require large hardware investments, have high costs, and have a limited number of monitoring objects, making them unsuitable for rapid and convenient on-site monitoring. The three-dimensional high-speed imaging system selects a portable, non-contact, low-cost optical monitoring system, realizing timely and effective detection of the operational status of GIS shell equipment in the power grid, while meeting the requirements of overall equipment portability and low cost. Attached Figure Description

[0027] Figure 1 This is a flowchart of the present invention; Figure 2 This is a detailed description of the fully connected process of the first neuron in the input layer of the MLP in this embodiment of the invention; Figure 3 This demonstrates the multilayer perceptron network structure used for feature extraction in an embodiment of the present invention; Figure 4 This is a schematic diagram of the C2C deformation comparison analysis results of an embodiment of the present invention; Figure 5 The results of the C2M deformation comparison analysis in the embodiments of the present invention; Figure 6The results of the deformation comparison analysis of M3C2 in the embodiment of the present invention; Figure 7 The results of deformation comparison analysis are shown in the embodiments of the present invention. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0029] like Figure 1 As shown in the figure, the GIS equipment deformation detection method based on point cloud registration described in this embodiment performs the following steps using a computer device: Example 1 This embodiment provides a technical solution: a GIS equipment deformation detection method based on point cloud registration, comprising the following steps: S1: Data Acquisition Use a binocular camera to acquire 3D point cloud data from a GIS device; S2: Data Preprocessing Preprocessing the acquired 3D point cloud data, such as sampling and filtering, can significantly improve the quantity and quality of the point cloud, providing a foundation for subsequent point cloud analysis. S3: Point Cloud Registration A preliminary registration method based on a novel feature descriptor PN is proposed. The method uses a multilayer perceptron (MLP) to extract features from the collected point cloud data to obtain a novel feature descriptor PN. Then, the RANSAC algorithm is used for feature matching to establish an approximate correspondence and achieve preliminary alignment of two point clouds. By combining the KD tree nearest neighbor search method to accelerate the ICP algorithm, accurate registration of two point clouds is achieved. In this embodiment, the two point clouds are the pre-processed point cloud data (i.e., the measured point cloud) and the ideal point cloud, respectively.

[0030] S4: Deformation Detection Based on the point cloud registration results, the displacement distance between the measured point cloud (i.e., the above-mentioned 3D point cloud data) and the ideal point cloud is calculated. This displacement distance is used as the main parameter for identifying the deformed area. The deformation threshold X is calculated using the deformation threshold formula. By comparing with displacement distance, all deformation points are identified and integrated to determine the deformation area of ​​the GIS equipment; S5: Model Predictive Analysis By employing a lightweight deep learning model, and through detailed analysis and risk assessment of deformable regions, the deep learning model can predict the degree of risk that these regions may face in the future.

[0031] In step S2, the point cloud data acquired by the binocular camera suffers from an excessive number of points and a high density of points in near-field data, making it unsuitable for direct deformation detection. Therefore, the acquired point cloud data needs to be processed. A farthest point sampling operation is performed on the point cloud data. A point is randomly selected from the point cloud as the initial point and added to the sampling point set. Then, the farthest point is selected and added to the sampling point set, continuing until the required number of sampling points is reached.

[0032] In step S2, the acquired point cloud data contains noise and discrete points, making it unsuitable for direct deformation detection. Therefore, the acquired point cloud data needs processing. Calculating the average distance and standard deviation between each sampling point and its neighborhood points, and removing points whose distance exceeds a threshold, effectively removes isolated and noise points from the data. For each point in the point cloud, calculating its weighted average with its neighborhood points, with weights determined by a Gaussian filter, and using the calculated value as the new value for that point, effectively smooths the point cloud data and preserves more point cloud details.

[0033] In step S3, the point cloud registration process includes two main stages: preliminary registration and precise registration. Preliminary registration refers to matching two point clouds when their initial poses are completely unknown, with the aim of providing a good relative pose for precise registration. Preliminary registration minimizes the difference between corresponding points in the two point clouds when their relative poses are approximately the same.

[0034] In step S3, the process of obtaining the novel feature descriptor PN includes: operating on the preprocessed point cloud data, where these sampling points represent the global information of the entire point cloud; for these sampling points, using a spherical distance metric to find the neighborhood points corresponding to each sampling point, which are used for subsequent feature extraction operations; wherein, the spherical distance metric is the Euclidean distance from a point to other points; for each sampling point, performing a multilayer perceptron operation with its neighborhood points to convert its point cloud data into a 128-dimensional feature vector, which is used to represent the local features of the sampling point.

[0035] More specifically, the principle of the RANSAC algorithm is as follows: RANSAC (Random Sample Consensus) is an iterative method that can effectively extract valid data points that conform to the model from a dataset containing a large number of outliers, thereby obtaining reliable model parameter estimates.

[0036] More specifically, the basic steps of the RANSAC algorithm are as follows: Randomly select a subset of point pairs and calculate the initial transformation matrix; use the transformation matrix to perform a preliminary transformation on the point cloud and evaluate it against the ideal point cloud to observe whether it meets the number of points requirement; repeat the above steps until the consistency is maximized or the expected number of iterations is reached, and output the best model.

[0037] More specifically, the principle of the ICP algorithm includes: The ICP (Iterative Closest Point) algorithm is a classic algorithm for point cloud registration. Its basic idea is to iteratively find the nearest neighbor pair between two sets of point clouds and calculate the optimal rigid transformation so that one point cloud can be aligned with the other point cloud as much as possible.

[0038] More specifically, the basic steps of the KD-tree nearest neighbor search method include: A KD-tree (K-Dimensional Tree) is a spatial partitioning data structure used for efficient nearest neighbor search. When searching a point cloud set, it quickly locates neighboring points by constructing an index structure. Before point cloud registration, the point cloud to be registered and the ideal point cloud are spatially partitioned, and a kd-tree structure is constructed. This method can reduce the time complexity of subsequent nearest point searches from linear to logarithmic levels, significantly improving the computational speed of the ICP algorithm, especially on large-scale point cloud datasets.

[0039] In step S5, the deformation threshold calculation process includes: the threshold calculation method is related to the registration accuracy index and the quality of the acquired point cloud. The deformation threshold is determined by collecting the point cloud quality and registration accuracy, which is used to determine whether the device should analyze the deformation.

[0040] Example 2 This embodiment proposes a GIS deformation detection method for electrical equipment based on point cloud registration, such as... Figure 1 As shown, it includes the following steps: I. Data Collection The system uses a binocular camera to collect 3D point cloud data from GIS equipment, along with corresponding image data and real-time temperature information. A ZED2i binocular camera is used with the following parameters: Table 1 ZED 2i Camera Parameters

[0041] II. Data Processing First, the 3D point cloud data undergoes Farthest Point Sampling (FPS). A point is randomly selected from the point cloud as the initial point and added to the sampling point set. Then, the farthest point is selected and added to the sampling point set until the required number of sampling points is reached. This simplifies the 3D point cloud data while preserving its overall structure. The simplified point cloud is then filtered to remove discrete points. The average distance and standard deviation between each sampling point and its neighborhood are calculated, and points exceeding a threshold are removed, effectively eliminating isolated and noisy points. Further filtering is performed: for each point in the point cloud, a weighted average of its distance to its neighborhood is calculated, with weights determined by a Gaussian filter. This calculated value is used as the new value for that point, effectively smoothing the point cloud data and preserving more point cloud details. The farthest point is iteratively selected from the point cloud using the following formula:

[0042] in For the selected point set, This is the original point cloud.

[0043] III. Point Cloud Registration This embodiment proposes a preliminary registration method based on a novel feature descriptor PN, which achieves approximate registration of an ideal point cloud and a measured point cloud when the two point clouds are not completely identical. First, the two point clouds are input into a multilayer perceptron (MLP) for feature extraction, transforming the point cloud data into a 128-dimensional feature vector, which is called the novel feature descriptor PN. The MLP consists of multiple layers, each containing several neurons, with fully connected layers. For example, a MLP structure of [8, 16] indicates two hidden layers with 8 and 16 neurons respectively, and each neuron is fully connected to all neurons in the layer above it. Figure 2 This provides a detailed description of the fully connected process of the first neuron in the input layer of an MLP. Figure 3 This paper demonstrates the multilayer perceptron network structure used for feature extraction in this invention. Point-pair matching relationships are constructed as a second-order compatible graph (SOG). The Bron-Kerbosch algorithm is used to search for maximal cliques in the graph, selecting corresponding point sets with geometric consistency. The optimal transformation matrix is ​​solved by maximizing the objective function.

[0044] in The weights are the correspondence weights determined by the principal eigenvectors of the SOG weight matrix.

[0045] Then, the feature descriptor is used for RANSAC algorithm matching to establish an approximate correspondence and achieve preliminary alignment of the two point clouds. To further improve registration accuracy, the ICP algorithm is accelerated by combining the KD tree nearest neighbor search method, which registers the two point clouds to be detected into the same coordinate system, reduces the stitching error between the point clouds, and completes the accurate registration of the two point clouds.

[0046] IV. Algorithm Evaluation This embodiment uses two registration evaluation metrics, RMSE and overlap rate, to assess the accuracy of the proposed point cloud registration algorithm. Ten sets of point cloud data obtained through on-site power grid photography were used as experimental data. Finally, the performance of the algorithm was evaluated by the average of the evaluation metrics from the ten experiments.

[0047] V. Deformation Detection The deformation threshold is calculated using the deformation threshold calculation formula. This value, as the main parameter for identifying deformed points, is usually related to factors such as registration accuracy indicators, overlap rate, RMSE, and point cloud quality. Point cloud data quality refers to the number of tiny noise points that may be generated when data is collected under high temperature and extreme light conditions. These noise points significantly affect data quality and cannot be removed by traditional filtering methods. To ensure detection accuracy, a relatively large deformation threshold needs to be set. The deformation threshold calculation formula is as follows:

[0048] in, The minimum deformation threshold for determining whether deformation has occurred; A represents the overlap rate after accurate registration; A represents the point cloud data quality, determined by real-time temperature. 1. 2 represents the ratio factor between the two factors; Other influencing factors were directly derived by analyzing the stability of the shooting equipment.

[0049] The displacement distance and deformation threshold between corresponding points in the point cloud are obtained by calculation. And compare them. If the displacement distance of a certain point exceeds the threshold... If a point is found to be deformed, it is marked as a deformation point. This detection process is repeated for all points to filter out all deformable points, which are then aggregated to form a clearly defined deformation region. If a single deformable point is found, it may be due to noise in the point cloud acquisition; this point is removed. The displacement distance of each deformable point is recorded; this data will provide support for subsequent shell failure prediction algorithms.

[0050] To more accurately quantify local deformation and suppress noise, this invention proposes a local Hausdorff distance calculation method. The mathematical expression of this method is as follows: For points in the point cloud to be measured... Find the corresponding point with the closest Euclidean distance in the ideal point cloud. and with Centered on a preset radius Perform a ball query to obtain its neighborhood point set. .point Local deformation By calculating its distance to the set of neighboring points The average Euclidean distance of all points is obtained, and the calculation formula is:

[0051] in, This represents the number of points within the neighborhood point set. Compared to the traditional global Hausdorff distance, this method effectively suppresses the interference of outliers through local averaging and is more robust to noise.

[0052] To eliminate abrupt changes in deformation values ​​caused by data acquisition noise or local registration errors, a high-order polynomial fitting method based on the moving least squares approach is used to smooth the deformation field in the spatial domain. For each point to be smoothed in the deformation region... Take it nearest neighbor Using Gaussian radial basis functions As weights. Smoothed deformation. Calculated using the following formula:

[0053] in, The basis functions of the quadratic polynomial are formed. This is the coefficient vector obtained through weighted least squares fitting. This operation outputs a continuous, stable, and physically more reasonable deformation field.

[0054] VI. Model Predictive Analysis This invention utilizes a lightweight multilayer perceptron to achieve risk prediction and analysis, with the network structure as follows: Figure 3As shown, this model inputs the detected deformation area, the total displacement of the deformation points, and the corresponding collected temperature information into a trained prediction model to obtain the final prediction result. The specific steps are as follows: First, the point cloud data of the deformation area is input into a three-layer perceptron [64,128,1024] for feature extraction until the feature space dimension becomes 1024; max pooling is performed on the feature space dimension to extract the global feature vector from the point cloud; simultaneously, temperature and displacement vectors are input independently, and these three vectors are merged; the merged vector is input into a multilayer perceptron [512,256,1] for further processing, and a binary classification result is output through the sigmoid activation function. This binary classification result indicates "deformation risk exists" or "deformation risk does not exist," used to predict potential future risks to the equipment, such as further deformation or damage. If the model predicts an abnormal risk, the system will issue a risk warning, prompting relevant personnel to conduct further inspections or take appropriate maintenance measures.

[0055] To comprehensively evaluate the performance of the point cloud comparison method proposed in this chapter in GIS equipment deformation monitoring, the method was implemented using Python programming, and three mainstream algorithms from the CloudCompare open-source platform—C2C, C2M, and M3C2—were selected as comparison objects. The experiment systematically compared and analyzed the performance of each method under complex deformation scenarios using three indicators: average deformation distance, maximum deformation distance, and algorithm execution time.

[0056] (1) C2C point cloud comparison method As shown in the deformation area visualization, the deformation calculated by the C2C method is too large and cannot accurately locate the specific deformation position of the GIS device. According to... Figure 4 The bar chart shows that the average deformation of the equipment is 34.6 mm, the maximum deformation is 52.1 mm, and the deformation is mainly concentrated in the range of 0 to 40 mm. Among them, there are 15,385 points with deformation between 0 and 20 mm, accounting for 50.8%, and 5,401 points with deformation between 20 mm and 40 mm, accounting for 17.9%.

[0057] (2) Comparison method of C2M and M3C2 point clouds Visualization based on deformation region Figure 5 (a) and Figure 6 As can be seen in (a), the distribution results of the M3C2 and C2M methods in the deformed region are highly similar. Therefore, the results of the two methods are combined for analysis in this experiment. Compared with the ideal point cloud, the two sides of the point cloud to be monitored show significant deformation, while the deformation in the middle region is less. The average deformation of the device is 16.2 mm, and the maximum deformation is 34.7 mm. Among them, the red deformation area is due to the lack of point cloud data during the data acquisition process. From the deformation bar chart... Figure 5(b) and Figure 6 As shown in (b), most of the deformation is concentrated between 0 and 60 mm. In the M3C2 results, the number of points between 0 mm and 15 mm accounts for 69.1%; the number of points between 15 mm and 30 mm accounts for 9.45%. Compared with the C2C method, the deformation point distribution of the C2M and M3C2 methods is different. The number of points between 0 mm and 15 mm is reduced, and the number of points between 15 mm and 30 mm is increased. Moreover, the deformation is mainly concentrated on both sides of the equipment, which is more consistent with the actual deformation situation.

[0058] (3) Improved point cloud comparison method This experiment uses an improved point cloud comparison method to perform deformation analysis on two periods of point cloud data from a GIS device, generating a visualization map and statistical bar chart of the deformation field in the expansion joint area, such as... Figure 7 As shown. From Figure 7 As shown in (a), the overall surface of the equipment is predominantly blue, with only a few areas exhibiting slight deformation. Only some localized areas show light green and yellow, indicating that the main structure of the equipment remains stable. To quantify the deformation distribution characteristics, the experiment performed interval statistics on the deformation and plotted a bar chart, as shown below. Figure 7 As shown in (b), the vast majority of deformations are concentrated between 0 mm and 0.025 mm, with an average deformation of 16.9 mm and a maximum deformation of 13.7 mm, which is within the normal fluctuation range of millimeter-level accuracy. There are some red areas at the edge of the equipment; for the range of 0.04 to 0.045 in the histogram, the analysis suggests that this phenomenon may be due to missing point clouds during data acquisition, rather than structural deformation.

[0059] Overall, the surface deformation of the equipment is within the normal range, and no significant deformation has been observed. The improved point cloud comparison method, by combining octree neighborhood search with local Hausdorff distance, can effectively display the deformation region of the equipment and generate a continuous deformation field through polynomial fitting, providing local deformation features that conform to the actual situation.

[0060] (4) Comparative analysis of point cloud comparison methods This experiment used the method of this invention and three classic point cloud comparison methods to calculate the deformation of GIS equipment, and the results were discussed and analyzed. Table 1 shows that the C2C method calculated an average deformation of 34.6 mm, which has a large error compared to the actual value, indicating that its application effect in deformation monitoring of GIS equipment is not ideal. The deformation calculation results of the other methods are basically consistent, preliminarily verifying the effectiveness of the monitoring results of the method of this invention.

[0061] Although the M3C2 method boasts high accuracy, its long computation time and high resource consumption result in low efficiency. As shown in Table 2, the M3C2 method takes 28 seconds, significantly longer than the method described in this invention. Especially in engineering applications requiring high-frequency or real-time monitoring, this low computational efficiency can become a bottleneck. In contrast, the method described in this invention maintains good accuracy while offering a more reasonable computation time, thus balancing monitoring precision and computational efficiency.

[0062] Table 2 Calculation results of point cloud comparison method

[0063] In summary, the point cloud comparison method proposed in this invention achieves a good balance between accuracy and computational efficiency. This method not only accurately captures the deformation trend of GIS equipment surfaces but also possesses high computational efficiency, meeting the real-time requirements of engineering applications and providing an efficient and reliable technical solution for GIS equipment status monitoring.

[0064] In another aspect, the present invention also discloses a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the method described above.

[0065] In another aspect, the present invention also discloses a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the method described above.

[0066] In another embodiment provided in this application, a computer program product containing instructions is also provided, which, when run on a computer, causes the computer to execute any of the point cloud registration-based GIS device deformation detection methods described in the above embodiments.

[0067] It is understood that the systems, devices, and storage media provided in the embodiments of the present invention correspond to the methods provided in the embodiments of the present invention, and the explanations, examples, and beneficial effects of the relevant content can be referred to the corresponding parts of the above methods.

[0068] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially as a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid state disk (SSD)).

[0069] It should be noted that in this invention, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0070] The various embodiments in this specification are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0071] 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 the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A GIS device deformation detection method based on point cloud registration, characterized in that, The method comprises the following steps: Data acquisition, using a binocular camera to obtain three-dimensional point cloud data of GIS equipment, and simultaneously obtaining auxiliary information such as shooting time and real-time temperature; Data preprocessing, farthest point sampling and filtering processing are performed on the collected three-dimensional point cloud data to obtain preprocessed point cloud data; Point cloud registration, a feature descriptor PN generated based on a multilayer perceptron is combined with an RANSAC R algorithm to realize preliminary registration, and an ICP algorithm accelerated by a KD tree neighbor search method is combined to complete accurate registration; Deformation detection, on the basis of the registration result of the point cloud, the displacement distance between the measured point cloud and the ideal point cloud is calculated, and a threshold value X is determined according to a deformation threshold value formula to determine the deformation points exceeding the threshold value and integrate the deformation region; Model prediction analysis, the detected deformation region is input into a lightweight prediction model, combined with temperature, displacement vector and point cloud feature vector, and a deformation risk level prediction result is output.

2. The point cloud registration based GIS device deformation detection method of claim 1, wherein: The farthest point sampling step comprises the following steps: a point in the point cloud is randomly selected as an initial point, which is added to a sampling point set, then the farthest point from the sampling point set is selected in turn to join the sampling point set, and the process is repeated until a predetermined number is reached. The farthest point is iteratively selected from the point cloud, and the formula is as follows: wherein is the selected point set, is the original point cloud.

3. The point cloud registration based GIS device deformation detection method of claim 1, wherein: The filtering processing step comprises the following steps: the average distance and the standard deviation of each point in the point cloud and the neighborhood points are calculated, the discrete points exceeding the threshold value are removed, and the point cloud is smoothed based on a Gaussian filter.

4. The point cloud registration based GIS device deformation detection method of claim 1, wherein: The preliminary registration specifically comprises the following steps: A spherical distance measurement method is used to find the neighborhood points of each sampling point; A 128-dimensional feature descriptor PN is generated by extracting a feature vector through a multilayer perceptron; The PN feature descriptor is combined with an RANSAC algorithm to establish a point pair relationship, and the point cloud is preliminarily aligned.

5. The GIS equipment deformation detection method based on point cloud registration according to claim 1, characterized in that: The RANSAC algorithm generates an initial transformation matrix by randomly selecting a point pair subset, iteratively evaluates consistency, and obtains the best model parameters.

6. The GIS equipment deformation detection method based on point cloud registration according to claim 1, characterized in that: The accurate registration adopts an ICP algorithm, iteratively calculates the nearest neighbor points of the measured point cloud and the ideal point cloud, and obtains the optimal rigid transformation matrix.

7. The point cloud registration based GIS device deformation detection method of claim 1, wherein: The ICP algorithm constructs an index structure through a KD tree neighbor search method before execution, and reduces the complexity of searching for the nearest point from O(n) to O(log n).

8. The point cloud registration based GIS device deformation detection method of claim 1, wherein: The calculation formula of the deformation threshold value X is as follows: wherein, is the minimum deformation threshold for determining whether deformation occurs; is the overlap rate after accurate registration; A is the point cloud data quality, determined by real-time temperature; 1、 2 is the proportion factor of the two factors; is other influencing factors, obtained by analyzing the stability of the shooting device.

9. The point cloud registration based GIS device deformation detection method of claim 1, wherein: When the displacement of a single point cloud exceeds the threshold value X, if it is an isolated point, it is determined as noise and removed. In order to more accurately quantify local deformation and suppress noise, a local Hausdorff distance calculation method is included, and the steps are as follows: For points in the point cloud to be tested Find the corresponding point with the closest Euclidean distance in the ideal point cloud. and with Centered on a preset radius Perform a ball query to obtain its neighborhood point set. ;point Local deformation By calculating its distance to the set of neighboring points The average Euclidean distance of all points is obtained, and the calculation formula is: wherein, represents the number of points within the neighborhood set; In order to eliminate the deformation value mutation caused by data acquisition noise or local registration error, a high-order polynomial fitting method based on the moving least square idea is used to perform spatial domain smoothing processing on the deformation field. For each point to be smoothed in the deformation region , take its nearest neighbors , use a Gaussian radial basis function as the weight; the smoothed deformation variable is computed by the following equation: wherein is a quadratic polynomial basis function of the form is a vector of coefficients obtained by weighted least squares fitting.

10. The point cloud registration based GIS device deformation detection method of claim 1, wherein: The prediction model is a three-layer multilayer perceptron, the feature space dimension is expanded to 1024 layer by layer, the global feature vector is extracted through maximum pooling, and then the temperature and displacement vector are combined, and the sigmoid function is used to output the binary classification prediction result.

Images