Large-size lightning arrester parameter measurement method based on mechanical arm collaborative three-dimensional reconstruction

Abstract

The invention discloses a lightning arrester parameter measurement method, which comprises the following steps: firstly, acquiring point cloud data of a power scene by using a depth camera, and outputting initial position attitude information of a lightning arrester; then controlling a mechanical arm to perform multi-view scanning on the lightning arrester according to the initial position attitude information to obtain multi-frame point cloud data of the target area; then performing preprocessing, feature extraction and registration on the obtained point cloud data to obtain registered point cloud data; and finally, performing three-dimensional reconstruction on the registered point cloud data to obtain a reconstructed lightning arrester three-dimensional model. And measuring related parameters of the lightning arrester according to the obtained three-dimensional model. According to the invention, the lightning arrester can be automatically positioned and three-dimensional reconstruction can be carried out, so that the detection efficiency and precision are greatly improved. In addition, the system has good robustness and can stably work in various complex environments. In the future, the technology is expected to be widely applied to maintenance and management of an electric power system.

Description

Technical Field

[0001] The present invention belongs to the field of power equipment status monitoring, and specifically relates to a three-dimensional reconstruction technology based on the collaboration of a robotic arm and a depth camera, which is used to measure the creepage distance of a lightning arrester and realize the intelligent operation and maintenance of the power system. Background Art

[0002] Lightning arresters are essential components in power systems, protecting electrical equipment from overvoltage surges. Creepage distance is a key indicator of their insulation performance, directly impacting their safety and reliability. Traditional creepage distance measurement methods typically rely on manual labor, which is not only time-consuming and labor-intensive, but also difficult to ensure accurate measurement when working with large linear arresters.

[0003] In recent years, with the advancement of sensor and robotics technologies, 3D reconstruction techniques based on depth cameras have been widely used in industrial inspection and automation. While these technologies can provide highly accurate 3D data, achieving accurate 3D reconstruction and creepage distance measurement for devices with complex structures and unique shapes, such as lightning arresters, remains challenging.

[0004] Existing 3D reconstruction technology mainly uses a fixed-position depth camera to obtain 3D data of the target object, but when dealing with complex geometric shapes such as lightning arresters with umbrella skirt structures (see Figure 2 ), a single perspective often fails to capture all necessary details. In addition, due to the diversity of lightning arrester installation locations, a fixed depth camera may not be able to flexibly adapt to various measurement scenarios.

[0005] Therefore, it is particularly important to develop a measurement device that can automatically adjust the shooting angle and achieve full coverage. Summary of the Invention

[0006] Traditional lightning arrester parameter detection relies on manual operation, which has problems such as low efficiency and difficulty in ensuring accuracy. Especially in the context of the continuous expansion of the power system scale and the increasingly complex operation and maintenance environment, traditional methods are increasingly difficult to meet actual needs. In order to solve these problems, the present invention proposes an automatic measurement method for lightning arrester parameters with a higher degree of intelligence and automation, and discloses an automatic measurement method for lightning arrester parameters based on deep learning and three-dimensional reconstruction technology. This method integrates deep learning, robotic arm control, three-dimensional reconstruction and other technologies, and can accurately and efficiently complete the parameter measurement task of the lightning arrester under complex backgrounds.

[0007] A method for reconstructing a three-dimensional model of a lightning arrester comprises: first, using a depth camera to acquire point cloud data of an electric power scene and outputting initial position and posture information of the lightning arrester; then, controlling a robotic arm to perform multi-perspective scanning of the lightning arrester based on the initial position and posture information to acquire multi-frame point cloud data of a target area; then, preprocessing, feature extraction, and registration are performed on the obtained point cloud data to obtain registered point cloud data; and finally, performing three-dimensional reconstruction on the registered point cloud data to obtain a three-dimensional model of the lightning arrester.

[0008] Depending on the parameter calculation requirements, the above 3D model can be a partial surface model that includes all the insulator shed structures of the arrester to be tested, or a complete 3D model of the arrester. For example, calculating creepage distance can be completed using a partial surface 3D model. Of course, directly obtaining the entire 3D model of the arrester can also achieve the same result.

[0009] In the step of determining the initial position and posture information of the lightning arrester, if it is determined that the point cloud data does not contain a lightning arrester, the point cloud data of the power scene can be reacquired by adjusting the shooting angle of the depth camera, and this process can be repeated until the target lightning arrester is obtained, and the lightning arrester positioning result, i.e., the initial position and posture information of the lightning arrester, is finally output. When there are multiple lightning arresters in the point cloud data obtained for the power scene, the initial position and posture information of the multiple lightning arresters is output simultaneously, and subsequent operations are performed on each lightning arrester separately, ultimately obtaining a surface 3D model or a complete 3D model of each lightning arrester.

[0010] Furthermore, a pre-trained PointNet++ network model can be used to accurately locate the arrester in the scene and output its initial position and posture information. Based on this initial position and posture information, a robotic arm can be controlled to perform multi-view scans of the arrester to obtain multi-frame point cloud data of the target area. Using the PointNet++ network model can effectively filter out background interference, accurately locate the arrester in the scene, and output its initial position and posture information.

[0011] In the multi-view point cloud acquisition step, in order to obtain complete three-dimensional information of the lightning arrester, we use a robotic arm equipped with a depth camera to scan it from multiple angles. More specifically, the robotic arm is controlled to drive the depth camera to perform multi-view scans of the lightning arrester according to a preset trajectory or visual servoing method, and obtain multi-frame point cloud data of the target area to cover the surface of the lightning arrester to the greatest extent. In the actual modeling process, considering the size differences of the lightning arrester, we adopt a targeted strategy: for ordinary lightning arresters (or smaller lightning arresters), only two angles need to be collected (taking the vertical setting of the lightning arrester center axis as an example, generally the upper and lower axis side angles, such as the angles of 30 to 50 degrees and 120 to 140 degrees with the vertical) to effectively cover its surface; for longer lightning arresters, it is necessary to increase the scanning angle to ensure that more comprehensive point cloud information is obtained.

[0012] Furthermore, the preprocessing includes operations such as denoising, filtering, and outlier removal. Preprocessing can improve the quality of point cloud data and lay the foundation for subsequent registration and reconstruction.

[0013] Furthermore, SHOT (Signature of Histogram s of Orientations) feature descriptors are extracted from the pre-processed point cloud data to complete the feature extraction.

[0014] SHOT features can effectively express local geometric information and are robust to noise and occlusion. The multi-frame point cloud is then registered using the feature-based Fast Global Registration (FGR) algorithm, unifying point cloud data from different perspectives into the same coordinate system and stitching them together into a complete arrester point cloud model.

[0015] Furthermore, a Poisson surface reconstruction algorithm is used to perform the 3D reconstruction on the registered point cloud data. This reconstruction generates a high-precision, high-resolution 3D model of the arrester. The Poisson reconstruction algorithm effectively handles noise and missing data, reconstructing a smooth, complete model surface.

[0016] A method for reconstructing a three-dimensional model of a lightning arrester comprises the following steps:

[0017] (1) Arrester positioning: Use a depth camera to obtain point cloud data of the power scene and input it into the pre-trained PointNet++ network model to obtain the initial position and posture information of the arrester in the scene;

[0018] (2) Multi-view point cloud acquisition: Based on the initial position and posture information obtained, the robotic arm is controlled to drive the depth camera to perform multi-view scanning of the arrester to obtain multi-frame point cloud data covering the surface of the arrester;

[0019] (3) Point cloud preprocessing: preprocessing the collected multi-frame point cloud data, including denoising, filtering, and outlier removal operations;

[0020] (4) Feature extraction and registration: Extract the SHOT feature descriptor of the preprocessed point cloud data; then use the FGR algorithm to register the multi-frame point cloud, unify the point cloud data from different perspectives into the same coordinate system, and splice them into a complete arrester (surface) point cloud model;

[0021] (5) Three-dimensional reconstruction: The Poisson surface reconstruction algorithm is used to perform three-dimensional reconstruction on the registered point cloud data to generate a three-dimensional model of the arrester (surface) with a smooth and complete model surface.

[0022] A method for automatically calculating relevant parameters of a lightning arrester adopts any of the above-mentioned lightning arrester three-dimensional model reconstruction methods to construct a three-dimensional model of the lightning arrester to be calculated, and calculates relevant parameters of the lightning arrester based on the three-dimensional model.

[0023] The relevant parameters include but are not limited to the arrester's creepage distance, structural dimensions, etc. Taking creepage distance detection as an example, the creepage distance measurement path and calculation method are determined on the reconstructed 3D arrester model according to power industry standards, and the arrester's creepage distance is automatically calculated.

[0024] The present invention perfectly solves the problems existing in the prior art by using a method in which a robotic arm carries a depth camera. By dynamically adjusting the position and posture of the robotic arm, it can ensure that the depth camera can capture the three-dimensional information of the lightning arrester, especially the large-size linear lightning arrester, from the optimal angle, thereby realizing the precise three-dimensional reconstruction of the complex structure of the lightning arrester and the accurate measurement of the creepage distance.

[0025] This invention provides a novel method for reconstructing a lightning arrester's three-dimensional model and detecting its parameters. This method can automatically locate the arrester and perform three-dimensional reconstruction, significantly improving detection efficiency and accuracy. Furthermore, the system exhibits excellent robustness and can operate stably in various complex environments. In the future, this technology is expected to be widely used in the maintenance and management of power systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 Schematic diagram of the flow of the three-dimensional reconstruction steps of the lightning arrester under complex background based on this method.

[0027] Figure 2 The figure is a photo of a common lightning arrester in the prior art.

[0028] Figure 3 It is a picture collected by a depth camera for a certain scene.

[0029] Figure 4is the pre-processed point cloud data.

[0030] Figure 5 The image after registration.

[0031] Figure 6 This is the image obtained after 3D reconstruction. DETAILED DESCRIPTION

[0032] In order to describe the present invention in more detail, the technical solution of the present invention is described in detail below in conjunction with specific embodiments, but the scope of protection of the present invention is not limited thereto.

[0033] (1) The depth camera acquires point cloud data.

[0034] When using a depth camera to capture a scene containing a specific object (such as a lightning arrester), the depth camera can not only capture a standard RGB image, but also simultaneously obtain the depth information corresponding to each pixel. This information can be combined to generate a 3D point cloud dataset for subsequent data processing and analysis. Assuming that the average distance from the camera to the target object is d, the point cloud dataset obtained by the depth camera can be expressed as P = {p1, p2, ..., p n}, where each point pi=(x i ,y i ,z i ) represents the actual three-dimensional coordinate corresponding to the pixel in space. Here x i ,y i are the horizontal and vertical positions, and zi is the depth value of the point relative to the camera.

[0035] In this way, the present invention can obtain not only the position of each point on the image plane, but also its three-dimensional position in the real world, which is crucial for building accurate 3D models of objects such as lightning arresters.

[0036] Figure 3 This is a point cloud data map containing a lightning arrester collected in this embodiment.

[0037] (2) Automatic lightning arrester positioning network based on PointNet++

[0038] To accurately identify lightning arresters in complex backgrounds, this invention abandons traditional recognition methods that rely on manual identification or two-dimensional images. Instead, it innovatively employs deep learning technology, using the PointNet++ network model to directly process point cloud data captured by a depth camera to automatically locate the arrester. As a powerful point cloud processing tool, the PointNet++ network model effectively learns the spatial geometric characteristics of point clouds and is highly robust to point cloud sparsity and noise. This allows for accurate identification of target arresters even in complex power scenarios.

[0039] Specifically, firstly, the point cloud data P obtained by the depth camera is converted into {p1, p2, ..., p n} is input to the PointNet++ network. The network learns the global and local features of the point cloud through multi-level feature extraction and outputs the probability value s of each point belonging to the arrester i By setting a threshold, the point cloud set P_{target}={p_i|s_i>t} belonging to the arrester is selected, and its center of mass coordinates are calculated as the initial position of the arrester, providing target guidance for the subsequent precise control of the robotic arm.

[0040] Through this step, the initial position and posture information of the corresponding lightning arrester can be obtained.

[0041] (3) Multi-view point cloud acquisition based on robotic arm collaboration

[0042] To obtain complete three-dimensional information about the arrester, we used a robotic arm equipped with a depth camera to scan it from multiple angles. Taking into account the size differences of arresters, we adopted a targeted strategy: for ordinary arresters, only two angles (i.e., the upper and lower axial angles when the arrester is set vertically) are needed to effectively cover the surface of the arrester; for longer arresters, the scanning angle needs to be increased to ensure more comprehensive point cloud information. Finally, we stitched and fused multiple frames of point cloud data to construct a complete three-dimensional model of the arrester.

[0043] (4) Point cloud data preprocessing

[0044] In order to further optimize the point cloud data processing process, for each point cloud dataset P = {p1, p2, ..., p n}, perform the following preprocessing steps:

[0045] Denoising: Use statistical filtering to remove outliers. If the point p in the point cloud i If the number of neighboring points of is lower than the threshold, it is considered as a noise point and removed. This step helps improve the accuracy of subsequent processing.

[0046] Smoothing: Use Gaussian filtering or bilateral filtering to smooth the point cloud and reduce the impact of noise. Smoothing can improve the quality of the point cloud and make subsequent feature extraction more accurate.

[0047] The above preprocessing steps can improve the quality of the point cloud data, providing more accurate basic data for subsequent feature extraction and point cloud registration. These optimization measures are crucial for building a complete and accurate 3D model of the arrester.

[0048] like Figure 4 As shown, these are the three-dimensional point cloud images of the two visions after preprocessing.

[0049] (5) Feature extraction and point cloud registration

[0050] To achieve accurate registration of multi-view arrester point cloud data, this paper first uses the SHOT feature descriptor to extract the local geometric features of the point cloud. Specifically, for each point in the point cloud, a local reference frame is defined to ensure rotation invariance. The covariance matrix M is constructed using the k neighboring points around the feature point. The covariance matrix M is as follows.

[0051]

[0052] Among them, p i is the i-th (i=0,1,2....k) neighborhood point, p represents the current coordinate point, R is the radius of the local reference system, d i Represents the distance from the i-th point to point p.

[0053] Next, it divides the spherical support region within this local system into 32 regions, each with 11 subregions. It represents each subregion with a histogram, creating a descriptor with 352 elements. To construct this descriptor, it calculates the cosine of the angle between each point within each partition and the coordinate origin (feature point) and stores this information in the corresponding histogram bin. It produces a preliminary histogram description of the feature points. It uses a quadratic linear interpolation method to mitigate edge effects and smooth descriptor changes. This process captures the angular relationship between the point and the feature point in each partition of the local coordinate system. It then interpolates and normalizes the resulting geometrically rich histogram descriptor to finally obtain the SHOT feature descriptor for the point.

[0054] Then the FGR algorithm is used for point cloud registration. The algorithm first uses the SHOT feature descriptor to establish the corresponding point pairs between the two frames of point clouds, and then constructs the objective function.

[0055] The algorithm generates a relationship set based on quaternion constraints and verifies the point cloud relationship pairs generated by the feature descriptor. First, calculate the relationship pairs K1 generated by the feature descriptor, then check whether the two points in each pair of relationships are the closest points to each other, and put the relationship pairs that meet the conditions into K2. Finally, for any four relationship pairs in k2 (p1, q1), (p2, q2), (p3, q3) and (p4, q4), check the constraints by the following formula and put the relationship k3 that meets the conditions into the pair. Introduce the test constraint, get the satisfied relationship condition, and put k3 into the pair. The test constraint relationship is as follows:

[0056]

[0057] where p i is the i-th point of point cloud p, qi is the i-th point of point cloud q. When τ is 0.95, this constraint can verify whether the four relationship pairs are compatible.

[0058] After generating the above relationship set, the following objective function can be established based on the construction of the relationship set:

[0059]

[0060] Where C is the translation vector, l p，q For the corresponding point pair, μ controls the range in which the residual has a significant impact on the target, generally starting from twice the maximum surface radius D and gradually decreasing as the optimization proceeds. The objective function optimization process is related to l p，q The partial derivative of the equation is set to zero, and the partial derivative of E(C, L) is minimized, as shown in the following formula.

[0061]

[0062] In order to solve l p，q , which is calculated as follows.

[0063]

[0064] Finally, the minimum value of the relationship set target can be found, that is, the optimal solution for registration can be found.

[0065] Figure 5 This is the point cloud data after registration.

[0066] (6) Poisson reconstruction and parameter calculation

[0067] On the basis of point cloud registration, the present invention fuses the point cloud data of all view angles to obtain a complete arrester point cloud model. Figure 6 To obtain a more detailed and complete 3D model of the arrester, this paper uses the Poisson surface reconstruction algorithm to reconstruct the point cloud data. Compared to other reconstruction algorithms, the Poisson reconstruction algorithm can better handle noise and missing data, generating a smoother and more detailed 3D model, providing a reliable data foundation for subsequent accurate parameter measurement.

[0068] Finally, based on the reconstructed 3D arrester model, the present invention can automatically calculate arrester parameters, such as creepage distance and structural dimensions, according to power industry standards. Compared with traditional manual measurement methods, this method greatly improves measurement efficiency and accuracy, providing technical support for the intelligent operation and maintenance of power equipment.

Claims

1. A method for measuring parameters of a lightning arrester, characterized in that: include: Firstly, a depth camera is used to obtain point cloud data of the power scene and output the initial position and posture information of the arrester. Then, according to the initial position and posture information, the robotic arm is controlled to perform multi-perspective scanning of the arrester to obtain multi-frame point cloud data of the target area. Then, the obtained point cloud data is preprocessed, feature extracted and registered to obtain the registered point cloud data. Finally, the registered point cloud data is three-dimensionally reconstructed to obtain a reconstructed three-dimensional model of the arrester. The relevant parameters of the arrester are measured based on the obtained three-dimensional model.

2. The method for measuring arrester parameters according to claim 1, characterized in that: The pre-trained PointNet++ network model is used to accurately locate the lightning arrester in the scene and output its initial position and posture information.

3. The method for measuring arrester parameters according to claim 1, characterized in that: The preprocessing includes denoising, filtering, and outlier removal operations.

4. The method for measuring arrester parameters according to claim 1, characterized in that: The SHOT feature descriptor of the preprocessed point cloud data is extracted to complete the feature extraction.

5. The method for measuring arrester parameters according to claim 1, characterized in that: The FGR algorithm is used for point cloud registration.

6. The method for measuring arrester parameters according to claim 1, characterized in that: The Poisson surface reconstruction algorithm is used to perform the three-dimensional reconstruction on the registered point cloud data.

7. The method for measuring arrester parameters according to claim 1, characterized in that: include: (1) Lightning arrester positioning: Use a depth camera to obtain point cloud data of the power scene and input it into the pre-trained PointNet++ network model to obtain the initial position and posture information of the lightning arrester in the scene; (2) Multi-view point cloud acquisition: Based on the initial position and posture information obtained, the robotic arm is controlled to drive the depth camera to perform multi-view scanning of the arrester to obtain multi-frame point cloud data covering the surface of the arrester; (3) Point cloud preprocessing: preprocess the collected multi-frame point cloud data, including denoising, filtering, and outlier removal operations; (4) Feature extraction and registration: Extract the SHOT feature descriptor of the preprocessed point cloud data; then use the FGR algorithm to register the multi-frame point cloud data, unify the point cloud data from different perspectives into the same coordinate system, and splice them into complete arrester point cloud data; (5) 3D reconstruction: The registered point cloud data is reconstructed in 3D using the Poisson surface reconstruction algorithm to generate a 3D model of the arrester surface with a smooth and complete model surface; (6) Use the obtained three-dimensional surface model to obtain the relevant parameters of the lightning arrester.

8. The method for measuring arrester parameters according to claim 1, characterized in that: The relevant parameters include the creepage distance and structural dimensions of the arrester.

Images