Transformer rack component identification method and device based on point cloud processing technology

Abstract

The invention discloses a transformer rack part identification method and device based on a point cloud processing technology, and the method comprises the steps: obtaining needed rack data through employing an unmanned plane carrying a laser radar, and carrying out the data preprocessing, including statistical filtering, farthest point downsampling, and CSF ground removal; improving a parameter selection part of the RANSAC straight line segmentation method, wherein the parameter selection part comprises straight line identification times, a straight line width threshold value, a point selection strategy and a reference straight line direction; performing RANSAC straight line fitting on the rack point cloud by using the selected parameters, and adding a result checking part to improve the recognition accuracy, including straight line gap checking and width adaptive adjustment; and constructing a transformer component segmentation model based on the improved PointNet + + network, wherein the transformer component segmentation model comprises an input data format modification module, a residual connection module and a transformer module. And segmenting the transformer part by using the model. According to the method, the integrity and accuracy of cross arm identification can be ensured, and meanwhile, the segmentation precision of the transformer rack part is improved by modifying the RANSAC straight line segmentation algorithm and the original PointNet + + network.

Description

Technical Field

[0001] The present invention relates to a component recognition method and device on transformer rack point cloud data, which utilizes the transformer rack point cloud and the geometric characteristics of the components to perform RANSAC straight line fitting and PointNet++ network segmentation to segment the components of the rack in the point cloud. Background Art

[0002] As a key hub for power supply and transmission in the power system, transformer racks bear the important responsibilities of voltage conversion and power distribution. Their stable operation is crucial to ensuring the safety and reliability of the power grid. In order to monitor the operating conditions of these devices, power companies usually take regular inspections to record and evaluate the working conditions of each substation. However, transformer rack inspections still rely heavily on manual inspections, which is inefficient and has high labor costs.

[0003] The use of point cloud segmentation technology when inspecting transformer racks can effectively improve inspection efficiency and reduce labor costs. Point cloud segmentation is to divide point clouds according to spatial, geometric and texture features, so that point clouds in the same divided area have similar features. However, the existing research on point cloud segmentation technology is more focused on the fields of unmanned driving or medical imaging. There is a lack of research on point cloud segmentation of power equipment, especially transformer racks, and the existing related point cloud segmentation technology is more focused on identifying power lines, clustering and matching substation power equipment to identify component models, etc. For example, a real-time power line flight device and method based on point cloud data (CN118819184A) uses RANSAC to segment power lines, but this method cannot correctly distinguish between the dense interference point cloud part and the component point cloud part when used on the rack. Similarly, a power grid inspection precision control method and system based on drones (CN118838412A) identifies power equipment by clustering, but because the components of the rack are connected to each other, the components cannot be correctly separated from each other by this method in point cloud data. These methods will have problems of low component segmentation accuracy or poor segmentation precision when applied to transformer bench inspection. Summary of the invention

[0004] The present invention aims to overcome the problems of low segmentation accuracy and poor segmentation precision in the case of transformer rack component segmentation in the prior art, and proposes a transformer rack component identification method based on point cloud processing technology.

[0005] The technical solution adopted by the present invention to solve the problems of the prior art is:

[0006] The first aspect of the present invention relates to a transformer rack component identification method based on point cloud processing technology, comprising the following steps:

[0007] 1. A transformer rack component identification method based on point cloud processing technology, characterized in that it includes the following steps:

[0008] S1: First, use a UAV carrying a laser radar to obtain the required bench data and perform data preprocessing, including: statistical filtering, farthest point downsampling, and CSF ground removal;

[0009] S2: Then improve the parameter selection part of the RANSAC line segmentation method, including: the number of line recognition times, line width threshold, point selection strategy and reference line direction;

[0010] S3: Then, the selected parameters are used to perform RANSAC straight line fitting on the bench point cloud, and a result inspection part is added to improve the recognition accuracy, including: straight line gap inspection and width adaptive adjustment;

[0011] S4: Finally, a transformer component segmentation model based on the improved PointNet++ network is constructed, including: input data format modification, residual connection module and transformer module. The model is used to segment transformer components;

[0012] Preferably, the step S1 specifically includes:

[0013] S1.1: As described in claim 1, using the laser radar of a drone to capture on-site point cloud data with a transformer rack, specifically, capturing the transformer rack in a flat area or an area with few buildings can significantly improve the quality of the point cloud data. At the same time, when obtaining drone data, selecting the shooting distance and removing point clouds that are too far away can also reduce the amount of data and increase the subsequent processing speed.

[0014] S1.2: The statistical filtering in the data preprocessing as described in claim 1, specifically, needs to remove outliers captured in the on-site rig point cloud data. The presence of outliers will reduce the subsequent segmentation efficiency and segmentation quality, and statistical filtering removes outliers by averaging the distances to the surrounding point clouds to remove points that are too far away, thereby eliminating outliers.

[0015] S1.3: The farthest point downsampling in the data preprocessing as described in claim 1, specifically, in order to improve the efficiency of subsequent segmentation, it is necessary to downsample the filtered point cloud. The farthest point downsampling can keep the overall framework of the point cloud unchanged on the basis of reducing the point cloud density, thereby reducing the impact of downsampling on the accuracy of component segmentation.

[0016] S1.4: The CSF ground removal in the data preprocessing as claimed in claim 1, specifically, using a cloth simulation algorithm (CSF) to remove the ground from the downsampled point cloud, and distinguishing between ground and non-ground points by simulating the drooping process of a piece of cloth on the point cloud data; specifically, the change in the height of the cloth at time t+1 can be expressed as:

[0017]

[0018] Among them, z(x,y,t) represents the height of the cloth at the (x,y) position, g is the gravitational acceleration, and the curvature terms and represent the degree of bending of the cloth; after a series of iterative calculations, the cloth gradually sags and touches the ground point, thereby distinguishing the ground point from other points.

[0019] Preferably, the step S2 specifically includes:

[0020] S2.1: the number of recognized straight lines in the improved parameters as described in claim 1, specifically, after obtaining the point cloud data in S1, determining the specific number of segmented gantry crossarms, the number is selected according to the construction time and the construction standard at that time, and the number corresponds to the number of segmentation results corresponding to the RANSAC segmentation;

[0021] S2.2: The straight line width threshold in the improved parameter as claimed in claim 1, specifically, on the basis of determining the number, setting the number of iterations to a corresponding number, and making a distinction between the width of the components when dividing the transformer crossarm and other crossarms; if it is identified as a transformer crossarm, the component width threshold is increased, otherwise the normal width threshold is used;

[0022] S2.3: The point selection strategy in the improved parameters as described in claim 1, specifically, the point selection range during identification will be manually limited to the point cloud that meets the expectations of the crossarm. Specifically, identifying a straight line requires at least two points as iteration points. Assume that P has been selected. 1 (x 1 ,y 1 ,z 1 ) is the first point, and P is to be selected 2 (x 2 ,y 2 ,z 2 ) as the second point. Since the straight line where the gantry crossarm is located is basically parallel to the plane where the ground is located, you can set z when selecting the point. 2 -z 1 ≤|α|, where α is an empirical parameter representing the possible inclination of the crossarm. At the same time, due to the characteristics that the crossarm component must be attached to the pillar component, the pillar is segmented first, and the iteration point of the crossarm segmentation can be set within the pillar range to reduce the number of invalid iterations;

[0023] S2.4: The reference straight line direction in the improved parameters as described in claim 1, specifically, the straight line represented by the cross arm of the transformer rack is almost parallel to the horizontal ground, so before the iteration starts, the direction vector is set to [x, y, 0] as a reference, and the corresponding reference cross arm fitting straight line formula is set to l: y = kx + b, z = z 0 , where z 0 is the height coordinate corresponding to the fitting line. The corresponding straight line in the pillar fitting is perpendicular to the ground. At this time, the reference direction vector is [0,0,1], and the corresponding reference crossarm fitting straight line formula is l:x=x 0 ,y=y 0 , where (x 0 ,y 0 ) are the coordinates of the selected fitting points.

[0024] Preferably, the step S3 specifically includes:

[0025] S3.1: Using the parameters set in claim 3, perform RANSAC straight line fitting on the filtered and ground-removed gantry point cloud to obtain a preliminary recognition result;

[0026] S3.2: The straight line gap check in the result check as described in claim 1, specifically, if the segmentation component is a transformer stand support, the gap check part is skipped, and the fitting straight line formula of claim 3 is l:y=kx+b, and the reference point p 0 (x 0 ,y 0 ), iteration point p 1 (x 1 ,y 1 ), then:

[0027]

[0028] The reference point can be any point in the point cloud used for fitting. In this paper, the reference point is the centroid of the point cloud, that is, is the direction vector corresponding to the fitted straight line, is the corresponding vector from the iteration point to the reference point, and n 1 is the projection distance of the desired point on the straight line;

[0029] S3.3: Obtain the projection n corresponding to all iteration points 1 Then sort them in order of size, and then find the maximum value of the adjacent difference Δn max , and set the maximum tolerance threshold n base For comparison, if Δn max >n base, it is considered that there is too large a gap in the middle of the recognized straight line and the recognition result is abandoned; otherwise, it is considered that the gap check has passed;

[0030] S3.4: Adaptive adjustment of width in the result check as claimed in claim 1, specifically, after the maximum gap check, setting the distance threshold t used in claim 3 init is the initial threshold, halved as the comparison threshold, and t new =t init / 2, perform secondary segmentation, and record the number of internal points I under the two threshold conditions respectively init ,I new , if the number of interior points satisfies I new / I init >60%, the segmentation result is set as a straight line with half the size of the threshold, otherwise the initial threshold is continued as the standard.

[0031] Preferably, the step S4 specifically includes:

[0032] S4.1: The input data format of the transformer component segmentation model of the improved PointNet++ network as described in claim 1 is modified. Specifically, when inputting the network segmentation data, the input data is added with three-dimensional height after the traditional coordinates and normal vector coordinates, that is, (x, y, z, x n ,y n ,z n ,z,z,z), where x n ,y n ,z n Corresponding normal vector coordinates; Since the transformer component has obvious height characteristics, increasing the weight of the height coordinate during segmentation can effectively improve the segmentation effect of the transformer component;

[0033] S4.2: The residual connection module and transformer module of the transformer component segmentation model of the improved PointNet++ network as described in claim 1, specifically, the residual connection and transformer module are introduced on the basis of the original PointNet++ network to further improve the network segmentation performance, the residual connection refers to adding the feature representation MLP(x) obtained by the input x through the MLP to the input x itself, that is, the original feature representation, as the output, and the transformer module further improves the network segmentation accuracy by calculating the semantic relevance between different elements in the data sequence;

[0034] S4.3: Use the training set made from the bench data to train the network to obtain a training model, and use the improved input method to input the remaining bench point cloud data obtained in claim 3 to obtain the transformer segmentation result.

[0035] The second aspect of the present invention relates to a transformer rack component identification device based on point cloud processing technology, comprising a memory and one or more processors, wherein the memory stores executable code, and when the one or more processors execute the executable code, they are used to implement a transformer rack component identification method based on point cloud processing technology of the present invention.

[0036] A third aspect of the present invention relates to a computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the transformer rack component identification method based on point cloud processing technology of the present invention.

[0037] The present invention can solve the problem of low recognition accuracy and poor segmentation precision in the case of transformer rack component segmentation in the existing technology. It uses prior knowledge to adjust the parameters of the algorithm fitting process, and at the same time projects the point cloud onto a straight line to check the gap and compares the multi-threshold fitting data to ensure the integrity and accuracy of the cross arm identification. At the same time, the PointNet++ network is modified to improve the segmentation accuracy of transformer components. Finally, the goal of transformer rack component identification is achieved.

[0038] The present invention uses the rack point cloud data, RANSAC straight line fitting algorithm and PointNet++ network to identify all the components on the rack and obtain point cloud data with annotation information. The annotation information is used to distinguish the cross arm point clouds, the pillar point clouds, the transformer point clouds, and other point clouds in the rack point cloud, so as to realize the identification of the transformer rack components. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1a It is a schematic diagram of the process structure of the present invention, Figure 1b This is the improved PointNet++ network structure diagram of the present invention.

[0040] Figure 2 Schematic diagram of point selection range setting of the present invention, wherein blue represents the pillar point cloud that has been identified, and red represents other point clouds.

[0041] Figure 3 This is a schematic diagram of the point cloud gap check of the present invention. The blue represents the farthest projection of the left point cloud, and the red represents the interpolation of the nearest projection of the right point cloud and the farthest projection of the left point cloud, that is, the gap distance.

[0042] Figure 4a to Figure 4b Schematic diagram of the adaptive width threshold adjustment of the present invention, wherein: Figure 4a This is the error map when the width threshold is too high. Figure 4b This is the correct image after adjusting the threshold. Red represents the recognized point cloud of the gantry crossarm, and green represents other point clouds.

[0043] Figure 5It is a schematic diagram of the residual connection method of the present invention.

[0044] Figure 6 is the final segmentation effect diagram of the present invention, wherein: Figure 6 a is the effect diagram of the separation of the platform support and the cross arm. Figure 6 b is the effect diagram of the splitting of the bench transformer components.

[0045] Figure 7 It is a structural diagram of the device of the present invention. DETAILED DESCRIPTION

[0046] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0047] Example 1

[0048] This embodiment relates to a method for segmenting a transformer stand point cloud using a transformer stand component identification method based on a point cloud processing technology according to the present invention. Figure 1a The flowchart shown is specifically implemented, which includes the following steps:

[0049] Step S1: Use a UAV carrying a laser radar to obtain the required bench data and perform data preprocessing:

[0050] In this case, the transformer rack data was collected from some installed transformer racks in Xihu District, Hangzhou, all of which were collected on-site by DJI drones carrying lidar. A total of 14 valid transformer rack data were collected, corresponding to 14 complete transformer racks in different regions.

[0051] The judgment formula of outliers in statistical filtering is:

[0052] D max =μ+std*σ#(3)

[0053] In this case, the statistical filter selects 5 points around the statistics as outlier reference points, and the empirical parameter std is selected as 1 in this case. Then the statistical filter is performed using the data and formula.

[0054] The filtered gantry point cloud is further processed using the farthest point downsampling to reduce the data size. In this case, the farthest point sampling point is set to 20,000 points, that is, the original data is downsampled to a point cloud data with 20,000 points.

[0055] Then, the cloth simulation algorithm (CSF) is used to remove the ground from the downsampled point cloud, and the ground and non-ground points are distinguished by simulating the drooping process of a piece of cloth on the point cloud data. Specifically, the change in the height of the cloth at time t+1 can be expressed as:

[0056]

[0057] Among them, z(x,y,t) represents the height of the cloth at the (x,y) position, g is the acceleration of gravity, and the curvature terms and represent the curvature of the cloth; after a series of iterative calculations, the cloth gradually sags and touches the ground point, thereby distinguishing the ground point from other points. In this case, the number of CSF iterations is selected as 500 times, and the category threshold is adjusted to 1.0 to better remove the ground part that accounts for a large proportion of the collected data. The time interval parameter Δt is set to 1, and other data keep the default parameters unchanged.

[0058] Step S2: Improve the parameter selection part of the RANSAC line segmentation method:

[0059] S2.1: After obtaining the point cloud data in S2, determine the specific number of split crossarms. Here, take the 2016 three-phase transformer installation standard as an example, and select the number of crossarm splits as 5, corresponding to the transformer crossarm, lightning arrester crossarm, fuse crossarm, lead crossarm 1, and lead crossarm 2.

[0060] S2.2: The number of iterations is set to 1000 each time. In the 2023 standard for the installation of three-phase transformer stations, the maximum width of the transformer crossarm is about 0.2m, while the maximum width of the other four crossarms is about 0.1m, and the maximum width of the station support is about 0.3m. Therefore, the width thresholds are set to the corresponding width values ​​during segmentation to improve the segmentation accuracy.

[0061] S2.3: If Figure 2 The point selection range is set as shown in the figure. The point selection range during identification will be manually limited to the point cloud that meets the crossarm expectation. Specifically, identifying a straight line requires at least two points as iteration points. Assume that P has been selected. 1 (x 1 ,y 1 ,z 1 ) is the first point, and P is to be selected 2 (x 2 ,y 2 ,z 2 ) as the second point. Since the straight line where the gantry crossarm is located is basically parallel to the plane where the ground is located, you can set z when selecting the point. 2 -z 1≤|α|, where α is an empirical parameter representing the possible inclination of the crossarm. At the same time, due to the characteristics of the crossarm component that must be attached to the pillar component, the pillar is segmented first, and the iteration point of the crossarm segmentation can be set within the pillar range to reduce the number of invalid iterations. In this case, the empirical parameter α is set to 0.15, which means that when selecting points, the height difference between two points within 0.15m is considered to be the same crossarm.

[0062] S2.4: The straight line represented by the cross arm of the transformer rack is almost parallel to the horizontal ground, so the direction vectors are set to [x, y, 0] as reference before the iteration starts, and the corresponding reference cross arm fitting straight line formula is set to l:y=kx+b,z=z 0 , where z 0 is the height coordinate corresponding to the fitting line. The corresponding straight line in the pillar fitting is perpendicular to the ground. At this time, the reference direction vector is [0,0,1], and the corresponding reference crossarm fitting straight line formula is l:x=x 0 ,y=y 0 , where (x 0 ,y 0 ) are the coordinates of the selected fitting points.

[0063] Step S3: Use the selected parameters to perform RANSAC straight line fitting on the bench point cloud and add a result inspection part to improve the recognition accuracy:

[0064] S3.1: Use the parameters set in step S2 to perform RANSAC straight line fitting on the filtered and ground-removed gantry point cloud to obtain preliminary recognition results.

[0065] S3.2: If the segmented component is a transformer stand support, jump to S3.4, assume that the fitting straight line formula in step S3.1 is l:y=kx+b, and the reference point is p 0 (x 0 ,y 0 ), iteration point p 1 (x 1 ,y 1 ), then:

[0066]

[0067] The reference point can be any point in the point cloud used for fitting. In this paper, the reference point is the centroid of the point cloud, that is, is the direction vector corresponding to the fitted straight line, is the corresponding vector from the iteration point to the reference point, and n 1 is the projection distance of the desired point on the straight line; in this example, n baseThe maximum tolerance threshold is set to 0.1. That is, in the point cloud, if the maximum projection distance is more than 0.1m, it is considered that there is a large gap in the identified point cloud, and the crossarm recognition fails.

[0068] S3.3: Obtain the projection n corresponding to all iteration points 1 Then sort them in order of size, and then find the maximum value of the adjacent difference Δn max , and set the maximum tolerance threshold n base For comparison, if Δn max >n base , it is considered that there is too large a gap in the middle of the recognized straight line and the recognition result is abandoned; otherwise, it is considered that the gap check has passed.

[0069] S3.4: After the maximum gap check, set the distance threshold y used by S4 init is the initial threshold, halved as the comparison threshold, and t new =t init / 2, perform secondary segmentation, and record the number of internal points I under the two threshold conditions respectively init ,I new , if the number of interior points satisfies I new / I init >60%, the segmentation result is set to a straight line with half the size of the threshold, otherwise the initial threshold is used as the standard. In this case, the comparison thresholds of the pillars, transformer crossarms, and other crossarms are t new Set to 0.15m, 0.1m, and 0.05m respectively.

[0070] Step S4: Construct a transformer component segmentation model based on the improved PointNet++ network and use the model to segment the transformer components:

[0071] S4.1: When inputting network segmentation data, the input data is added with three-dimensional height after the traditional coordinates and normal vector coordinates, i.e. (x, y, z, x n ,y n ,z n ,z,z,z), where x n ,y n ,z n Corresponding normal vector coordinates; since the transformer component has obvious height characteristics, increasing the weight of the height coordinate during segmentation can effectively improve the segmentation effect of the transformer component.

[0072] S4.2: Based on the original PointNet++ network, residual connections and transformer modules are introduced to further improve the network segmentation performance. Figure 5The method shown introduces residual connection to change the SA layer of PointNet++, and the transformer module is connected between the SA layer and the FP layer. The final network structure is shown in Figure 1(b). Residual connection refers to adding the feature representation MLP(x) obtained by MLP of the input x to the original feature representation of the input x itself as the output. The transformer module further improves the network segmentation accuracy by calculating the semantic relevance between different elements in the data sequence.

[0073] In this case, the hardware device is a 6-core 3.4Ghz AMD CPU and an Nvidia3050Ti graphics card. The deep learning environment consists of Windows 11 system, python-3.7, pytorch-1.12, cu116, etc. The number of selected points in the original PointNet++ parameters is increased to 4096, and the number of iterations is increased to 100 times, and other parameters remain unchanged by default.

[0074] S4.3: Use the training set made from the bench data to train the network in S5.2 to obtain the training model. In this example, the training data uses the bench point cloud data that has been manually segmented and labeled. In order to match the method in this article, the bench in the training data will also artificially remove the ground, pillars, crossarms and other columnar parts. And use the input method of step S5.1 to input the remaining bench point cloud data obtained in step S4 to obtain the transformer segmentation result.

[0075] In this case, the final transformer component segmentation accuracy reached 97.4%, which is 0.9% higher than the 96.5% accuracy of the original PointNet++ network using the same parameters.

[0076] Step S5: Segment and label the transformer rack components according to the segmentation results:

[0077] According to the previous steps, the segmented cross arm component point cloud D can be obtained. 1 ,D 2 ,D 3 ,D 4 ,D 5 , and the point cloud D of the pillar component 6 ,D 7 , the crossarms are identified one by one according to the installation height of the crossarm components, that is, the component point clouds are numbered and sorted by comparing the height data of the point cloud centroid of the components. The centroid p can be obtained by calculating the point cloud centroid of the crossarm components. 1 ,p 2 ,p 3 ,p 4 ,p 5 , whose corresponding height is z 1 ,z2 ,z 3 ,z 4 ,z 5 . Assume that their heights have the following relationship: z 1 >z 2 >z 3 >z 4 >z 5 According to the component relationship, we can get D 1 is the point cloud of the lead cross arm 1, D 2 is the point cloud of the lead cross arm 2, D 3 is the point cloud of the fuse cross arm, D 4 is the point cloud of the arrester cross arm, D 5 is the transformer cross arm point cloud. Similarly, in the PointNet++ segmentation result, the point cloud data will increase from the input 9-dimensional to 10-dimensional, that is, (x, y, z, x n ,y n ,z n ,z,z,z,c). Where c = 0 or 1, 1 represents the identified transformer point cloud and 0 represents other point clouds. The transformer component point cloud is extracted through the label value given by the network.

[0078] The final segmentation result of the transformer rack column is as follows: Figure 6 As shown in (a), the transformer test bench transformer component segmentation results are as follows Figure 6 As shown in (b), the orange and red in Figure a correspond to the segmented point cloud of the stand support components, and the five straight lines from yellow to green in the middle correspond to the transformer crossarm, lightning arrester crossarm, fuse crossarm, lead crossarm 1 and lead crossarm 2 in yellow, light green, green, dark green and dark green, respectively, and the blue is the remaining part of the segmentation; correspondingly, the red in Figure b is the identified transformer component, and the blue is the remaining part of the transformer stand. It can be seen from the final segmentation result that the crossarm and support components are segmented accurately and with high precision.

[0079] In this case, there are a total of 112 components that need to be segmented in 14 test bench data, including 28 test bench pillars, 70 test bench cross arms and 14 test bench transformers. The point cloud segmentation accuracy of all components is higher than 94%. In this case, the method described in this patent realizes the segmentation task of the transformer test bench components.

[0080] The present invention utilizes the rack point cloud data, the RANSAC straight line fitting algorithm and the PointNet++ network to identify all components on the rack and obtain point cloud data with annotation information; the annotation information is used to distinguish the cross arm point clouds, the pillar point clouds, the transformer point cloud, and other point clouds in the rack point cloud, thereby realizing the identification of the transformer rack components.

[0081] Example 2

[0082] like Figure 7 This embodiment relates to a transformer rack component identification device based on point cloud processing technology, including a memory and one or more processors, wherein the memory stores executable code, and when the one or more processors execute the executable code, they are used to implement a transformer rack component identification method based on point cloud processing technology of the present invention.

[0083] Example 3

[0084] This embodiment relates to a computer-readable storage medium on which a program is stored. When the program is executed by a processor, the transformer rack component identification method based on point cloud processing technology of the present invention is implemented.

[0085] The embodiments of the present invention can be applied to transformer stand point cloud data collected by UAVs. The accuracy of the original algorithm can be improved by improving the iterative process by utilizing the geometric characteristics of the components, and the integrity and accuracy of the segmentation of columnar components can be further improved by checking the gap between straight line projections and selecting adaptive thresholds. At the same time, the modification of the PointNet++ network can also improve the segmentation accuracy of transformer component segmentation.

Claims

1. A transformer rack component identification method based on point cloud processing technology, characterized in that: The steps include: S1: First, use a UAV carrying a laser radar to obtain the required bench data and perform data preprocessing, including: statistical filtering, farthest point downsampling, and CSF ground removal; S2: Then improve the parameter selection part of the RANSAC line segmentation method, including: the number of line recognition times, line width threshold, point selection strategy and reference line direction; S3: Then, the selected parameters are used to perform RANSAC straight line fitting on the bench point cloud, and a result inspection part is added to improve the recognition accuracy, including: straight line gap inspection and width adaptive adjustment; S4: Finally, a transformer component segmentation model based on the improved PointNet++ network is constructed, including: input data format modification, residual connection module and transformer module. The model is used to segment transformer components.

2. A transformer rack component identification method based on point cloud processing technology as claimed in claim 1, characterized in that: The step S1 specifically includes: S1.1: As described in claim 1, using the laser radar of a drone to capture on-site point cloud data with a transformer rack, specifically, capturing the transformer rack in a flat area or an area with few buildings can significantly improve the quality of the point cloud data. At the same time, when obtaining drone data, selecting the shooting distance and removing point clouds that are too far away can also reduce the amount of data and increase the subsequent processing speed. S1.2: The statistical filtering in the data preprocessing as described in claim 1, specifically, needs to remove outliers captured in the on-site rig point cloud data. The presence of outliers will reduce the subsequent segmentation efficiency and segmentation quality, and statistical filtering removes outliers by averaging the distances to the surrounding point clouds to remove points that are too far away, thereby eliminating outliers. S1.3: The farthest point downsampling in the data preprocessing as described in claim 1, specifically, in order to improve the efficiency of subsequent segmentation, it is necessary to downsample the filtered point cloud. The farthest point downsampling can keep the overall framework of the point cloud unchanged on the basis of reducing the point cloud density, thereby reducing the impact of downsampling on the accuracy of component segmentation. S1.4: The CSF ground removal in the data preprocessing as claimed in claim 1, specifically, using a cloth simulation algorithm (CSF) to remove the ground from the downsampled point cloud, and distinguishing between ground and non-ground points by simulating the drooping process of a piece of cloth on the point cloud data; specifically, the change in the height of the cloth at time t+1 can be expressed as: Among them, z(x,y,t) represents the height of the cloth at the (x,y) position, g is the gravitational acceleration, and the curvature terms and represent the degree of bending of the cloth; after a series of iterative calculations, the cloth gradually sags and touches the ground point, thereby distinguishing the ground point from other points.

3. The transformer rack component identification method based on point cloud processing technology according to claim 1, characterized in that: The step S2 specifically includes: S2.1: the number of recognized straight lines in the improved parameters as described in claim 1, specifically, after obtaining the point cloud data in S1, determining the specific number of segmented gantry crossarms, the number is selected according to the construction time and the construction standard at that time, and the number corresponds to the number of segmentation results corresponding to the RANSAC segmentation; S2.2: The straight line width threshold in the improved parameter as claimed in claim 1, specifically, on the basis of determining the number, setting the number of iterations to a corresponding number, and making a distinction between the width of the components when dividing the transformer crossarm and other crossarms; if it is identified as a transformer crossarm, the component width threshold is increased, otherwise the normal width threshold is used; S2.3: The point selection strategy in the improved parameters as described in claim 1, specifically, the point selection range during identification will be manually limited to the point cloud that meets the expectations of the crossarm. Specifically, identifying a straight line requires at least two points as iteration points. Suppose P1 (x1, y1, z1) has been selected as the first point, and P2 (x2, y2, z2) will be selected as the second point. Since the straight line where the gantry crossarm is located is basically parallel to the plane where the ground is located, z2-z1≤|α| can be set when selecting points, where α is an empirical parameter representing the possible inclination of the crossarm. At the same time, due to the characteristics that the crossarm component must be attached to the pillar component, the pillar is segmented first, and the iteration point of the crossarm segmentation can be set within the pillar range to reduce the number of invalid iterations; S2.4: The reference straight line direction in the improved parameters as described in claim 1, specifically, the straight line represented by the cross arm of the transformer stand is almost parallel to the horizontal ground, so the direction vectors are set to [x, y, 0] as references before the iteration starts, and the corresponding reference cross arm fitting straight line formula is set to l: y = kx + b, z = z0, where z0 is the height coordinate corresponding to the fitting straight line. The pillar fitting corresponds to a straight line perpendicular to the ground, and the reference direction vector is [0, 0, 1] at this time, and the corresponding reference cross arm fitting straight line formula is l: x = x0, y = y0, where (x0, y0) is the coordinate of the selected fitting point.

4. The transformer rack component identification method based on point cloud processing technology according to claim 1, characterized in that: The step S3 specifically includes: S3.1: Using the parameters set in claim 3, perform RANSAC straight line fitting on the filtered and ground-removed gantry point cloud to obtain a preliminary recognition result; S3.2: The straight line gap check in the result check as claimed in claim 1, specifically, if the segmentation component is a transformer stand support, the gap check part is skipped, and the fitting straight line formula of claim 3 is l:y=kx+b, the reference point is p0(x0,y0), and the iteration point is p1(x1,y1), then: The reference point can be any point in the point cloud used for fitting. In this paper, the reference point is the centroid of the point cloud, that is, is the direction vector corresponding to the fitted straight line, is the corresponding vector from the iteration point to the reference point, and n1 is the projection distance of the desired point on the straight line; S3.3: After obtaining the projections n1 corresponding to all iteration points, sort them in order of size, and then obtain the maximum value Δn of the adjacent differences max , and set the maximum tolerance threshold n base For comparison, if Δn max >n base , it is considered that there is too large a gap in the middle of the recognized straight line and the recognition result is abandoned; otherwise, it is considered that the gap check has passed; S3.4: Adaptive adjustment of width in the result check as claimed in claim 1, specifically, after the maximum gap check, setting the distance threshold t used in claim 3 init is the initial threshold, halved as the comparison threshold, and t new =t init / 2, perform secondary segmentation, and record the number of internal points I under the two threshold conditions respectively init ,I new , if the number of interior points satisfies I new / I init If the value is >60%, the segmentation result is set as a straight line with half the size of the threshold value; otherwise, the initial threshold value is used as the standard.

5. The transformer rack component identification method based on point cloud processing technology according to claim 1, characterized in that: The step S4 specifically comprises: S4.1: The input data format of the transformer component segmentation model of the improved PointNet++ network as described in claim 1 is modified. Specifically, when inputting the network segmentation data, the input data is added with three-dimensional height after the traditional coordinates and normal vector coordinates, that is, (x, y, z, x n ,y n ,z n ,z,z,z), where x n ,y n ,z n Corresponding normal vector coordinates; Since the transformer component has obvious height characteristics, increasing the weight of the height coordinate during segmentation can effectively improve the segmentation effect of the transformer component; S4.2: The residual connection module and transformer module of the transformer component segmentation model of the improved PointNet++ network as described in claim 1, specifically, the residual connection and transformer module are introduced on the basis of the original PointNet++ network to further improve the network segmentation performance, the residual connection refers to adding the feature representation MLP(x) obtained by the input x through the MLP to the input x itself, that is, the original feature representation, as the output, and the transformer module further improves the network segmentation accuracy by calculating the semantic relevance between different elements in the data sequence; S4.3: Use the training set made from the bench data to train the network to obtain a training model, and use the improved input method to input the remaining bench point cloud data obtained in claim 3 to obtain the transformer segmentation result.

6. The transformer rack component identification method based on point cloud processing technology according to claim 1, characterized in that: By using the rack point cloud data, RANSAC straight line fitting algorithm and PointNet++ network, all components on the rack are identified and point cloud data with annotation information is obtained. The annotation information is used to distinguish the cross arm point clouds, pillar point clouds, transformer point clouds, and other point clouds in the rack point cloud to realize the identification of transformer rack components.

7. A transformer rack component identification device based on point cloud processing technology, characterized in that: It comprises a memory and one or more processors, wherein the memory stores executable codes, and when the one or more processors execute the executable codes, they are used to implement a transformer stand component identification method based on point cloud processing technology as described in any one of claims 1 to 6.

8. A computer-readable storage medium, characterized in that: A program is stored thereon, and when the program is executed by a processor, the transformer stand component identification method based on point cloud processing technology described in any one of claims 1 to 6 is implemented.

Images