Transformer substation three-dimensional map construction method based on point cloud registration

Abstract

The invention discloses a transformer substation three-dimensional map construction method based on point cloud registration, and the method comprises the steps: generating pseudo labels for unregistered point clouds through employing an iterative closest point ICP algorithm, carrying out the training and optimization through a deep neural network in combination with a small number of real labels, gradually improving the registration precision of the point clouds and the construction efficiency of a three-dimensional map, and improving the construction efficiency of the three-dimensional map. According to the method, the problem of scarcity of real label data caused by equipment shielding in a substation scene can be effectively solved, the construction quality and efficiency of the three-dimensional electronic map of the substation are remarkably improved, and the method has important practical application value.

Description

Technical Field

[0001] The present invention belongs to the technical field of power system operation and maintenance, and in particular relates to a method for constructing a three-dimensional map of a substation based on point cloud registration. Background Art

[0002] 3D laser scanning technology, with its high precision and high resolution, can fully capture the 3D point cloud data of the complex environment inside and outside the substation, providing a good foundation for the construction of 3D maps of substations. However, due to the wide variety and complex layout of equipment inside the substation, as well as the serious occlusion between equipment, the point cloud data obtained by laser scanning often has missing and irregular noise, which directly affects the accuracy and integrity of the 3D model.

[0003] The registration of point cloud data has become a key step in the 3D modeling process. Common registration algorithms such as the Iterative Closest Point (ICP) algorithm can complete the alignment of point clouds with high accuracy. However, when faced with large-scale unregistered data, this method has a high reliance on real labeled data, and it is often difficult to obtain accurate labeled data in substation scenarios. In addition, the complexity of substations means that there is still room for optimization in the computational efficiency and model accuracy of the traditional ICP algorithm and its variants. Some studies have introduced deep learning and semi-supervised methods to improve the registration effect by generating and optimizing pseudo-labels and reducing the reliance on real labels.

[0004] Therefore, it is necessary to design a substation three-dimensional map construction method based on point cloud registration to solve the above problems. Summary of the invention

[0005] The technical problem to be solved by the present invention is to provide a method for constructing a three-dimensional map of a substation based on point cloud registration. The method uses semi-supervised deep learning technology to continuously optimize the model by generating and selecting pseudo labels, aiming to solve the label scarcity problem caused by equipment occlusion in substation scenarios, improve the construction quality and efficiency of three-dimensional maps, and provide technical support for the intelligent management of power systems.

[0006] In order to achieve the above technical effects, the technical solution adopted by the present invention is: A method for constructing a three-dimensional map of a substation based on point cloud registration comprises the following steps: S1, using 3D laser scanning equipment to obtain multi-view point cloud data of equipment in the substation, including source point cloud and target point cloud ,get N The point cloud collection is ; A small amount of registered point cloud data is , a large amount of unregistered point cloud data is ; S2, using the iterative closest point algorithm ICP to generate pseudo labels for unregistered point cloud data to address the scarcity of real label data caused by equipment occlusion in the substation; S3, selects the generated pseudo labels and combines them with the real label point cloud data, and uses a semi-supervised deep neural network point cloud registration model to train and optimize the point cloud data to improve the accuracy of point cloud registration; S4, through the iterative optimization of the semi-supervised deep neural network point cloud registration model DCP, gradually improves the accuracy of pseudo-labels, thereby optimizing the construction effect of the three-dimensional map.

[0007] Preferably, in step S1, using a three-dimensional laser scanning device to obtain multi-view point cloud data of equipment in the substation includes the following process: Substation field staff conduct preliminary preparations by surveying the site, developing laser scanning plans and data collection operation procedures, and conducting 3D laser scanning of buildings and equipment in the station to obtain 3D point cloud data; Set up multiple points to perform all-round scanning of the target from different perspectives and positions, perform registration calculations on the point cloud data, and convert the 3D point cloud data in the independent coordinate systems of different points into a unified coordinate system; The three-dimensional point cloud model of the substation obtained through field and indoor processing is modeled in combination with its three-dimensional coordinates, color and size data. The small components are drawn in proportion using three-dimensional modeling software to construct a real and accurate three-dimensional electronic map of the substation.

[0008] Preferably, in step S2, using an iterative closest point algorithm to generate pseudo labels for unregistered point cloud data includes: Input of large unregistered point cloud datasets For the point cloud pairs in , set the threshold tau for algorithm termination; Calculate the closest point, for the i Point Cloud , under the condition of minimizing the neighboring distance, the following equation is solved and : ; in, is the rigid transformation matrix, R represents the rotation matrix predicted by the model, t represents the predicted translation vector; Calculate the rigid transformation. If the optimal correspondence between the source point cloud and the target point cloud is known, the optimal rigid transformation can be calculated as follows: ; in, To minimize the mean square error, and ; Target point cloud Medium distance The nearest point; Update the current point cloud and calculate the error between the current point cloud and the adjacent point cloud; determine whether the error is less than the threshold tau, if so, terminate the iteration, otherwise continue the iterative optimization calculation.

[0009] Preferably, in step S3, when training the deep neural network registration model, from the set without labels A subset of samples with pseudo labels is selected from the dataset to increase the number of labeled samples to reduce overfitting during model training. The selection of generated pseudo labels includes the following process: Train the DCP model on a small amount of registered point cloud data to obtain the initial model parameters ,parameter As the pre-trained parameters of the model SEMI-DCP, Indicates The values ​​of the model parameters at the iteration; Calculate the root mean square error when registering unlabeled point clouds ; Root mean square error For each point cloud pair, a tuple of five elements is formed ; according to Sort all tuples in ascending order of value; Select the first k pairs of point clouds and the corresponding pseudo labels together with the registered point cloud data Participate in the training of SEMI-DCP model; Finally, k pairs of pseudo-label samples are output .

[0010] Preferably, in step S4, during the iterative optimization of SEMI-DCP, the final model performance is improved by reducing the gap between the real label and the pseudo label. The specific process includes: During the pseudo-label update process, the rigid transformation predicted by the model trained in the previous iteration is used to initialize the initial pose of each point cloud pair; The rigid transformation and the corresponding r are calculated using the ICP algorithm. i , if the new r i Less than r stored in the quintuple i , then the quintuple R, t and r i Should be updated to new R, t and r i .

[0011] Preferably, a three-dimensional electronic map of the substation is drawn based on the registered three-dimensional point cloud data by three-dimensional modeling software to achieve three-dimensional modeling of equipment and environment.

[0012] The beneficial effects of the present invention are as follows: The present invention utilizes the iterative closest point (ICP) algorithm to generate pseudo labels for unregistered point clouds, and combines a small amount of real labels to perform training and optimization through a deep neural network, thereby gradually improving the registration accuracy of point clouds and the efficiency of three-dimensional map construction. The present invention can effectively solve the problem of scarcity of real label data caused by equipment occlusion in substation scenarios, significantly improve the construction quality and efficiency of three-dimensional electronic maps of substations, and has important practical application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 It is a flow chart of the semi-supervised deep neural network point cloud registration method of the present invention; Figure 2 This is a flowchart of the internal and external work of the substation of the present invention; Figure 3 It is a schematic diagram of laser radar scanning of the present invention; Figure 4 It is the SEMI-DCP iterative optimization process of the present invention; Figure 5 It is the transformer initial position and posture diagram obtained by the present invention through field processing; Figure 6 is the registration result diagram of the present invention; Figure 7 It is a graph of the effectiveness test results of the model of the present invention; Figure 8 It is a graph of the convergence test results of the model of the present invention. DETAILED DESCRIPTION

[0014] Embodiment 1: like Figure 1 As shown, a method for constructing a three-dimensional map of a substation based on point cloud registration includes the following steps: S1, using 3D laser scanning equipment to obtain multi-view point cloud data of equipment in the substation, including source point cloud and target point cloud ,get N The point cloud collection is ; A small amount of registered point cloud data is , a large amount of unregistered point cloud data is ; S2, using the iterative closest point algorithm ICP to generate pseudo labels for unregistered point cloud data to address the scarcity of real label data caused by equipment occlusion in the substation; S3, selects the generated pseudo labels and combines them with the real label point cloud data, and uses a semi-supervised deep neural network point cloud registration model to train and optimize the point cloud data to improve the accuracy of point cloud registration; S4, through the iterative optimization of the semi-supervised deep neural network point cloud registration model DCP, gradually improves the accuracy of pseudo-labels, thereby optimizing the construction effect of the three-dimensional map.

[0015] Preferably, in step S1, using a three-dimensional laser scanning device to obtain multi-view point cloud data of equipment in the substation includes the following process: Substation field staff conduct preliminary preparations by surveying the site, developing laser scanning plans and data collection operation procedures, and conducting 3D laser scanning of buildings and equipment in the station to obtain 3D point cloud data; Set up multiple points to perform all-round scanning of the target from different perspectives and positions, perform registration calculations on the point cloud data, and convert the 3D point cloud data in the independent coordinate systems of different points into a unified coordinate system; The three-dimensional point cloud model of the substation obtained through field and indoor processing is modeled in combination with its three-dimensional coordinates, color and size data. The small components are drawn in proportion using three-dimensional modeling software to construct a real and accurate three-dimensional electronic map of the substation.

[0016] Preferably, in step S2, using an iterative closest point algorithm to generate pseudo labels for unregistered point cloud data includes: Input of large unregistered point cloud datasets For the point cloud pairs in , set the threshold tau for algorithm termination; Calculate the closest point, for the i Point Cloud , under the condition of minimizing the neighboring distance, the following equation is solved and : ; in, is the rigid transformation matrix, R represents the rotation matrix predicted by the model, t represents the predicted translation vector; Calculate the rigid transformation. If the optimal correspondence between the source point cloud and the target point cloud is known, the optimal rigid transformation can be calculated as follows: ; in, To minimize the mean square error, and ; Target point cloud Medium distance The nearest point; Update the current point cloud and calculate the error between the current point cloud and the adjacent point cloud; determine whether the error is less than the threshold tau, if so, terminate the iteration, otherwise continue the iterative optimization calculation.

[0017] Preferably, in step S3, when training the deep neural network registration model, from the set without labels A subset of samples with pseudo labels is selected from the dataset to increase the number of labeled samples to reduce overfitting during model training. The selection of generated pseudo labels includes the following process: Train the DCP model on a small amount of registered point cloud data to obtain the initial model parameters ,parameter As the pre-trained parameters of the model SEMI-DCP, Indicates The values ​​of the model parameters at the iteration; Calculate the root mean square error when registering unlabeled point clouds ; Root mean square error For each point cloud pair, a tuple of five elements is formed ; according to Sort all tuples in ascending order of value; Select the first k pairs of point clouds and the corresponding pseudo labels together with the registered point cloud data Participate in the training of SEMI-DCP model; Finally, k pairs of pseudo-label samples are output .

[0018] Preferably, in step S4, during the iterative optimization of SEMI-DCP, the final model performance is improved by reducing the gap between the real label and the pseudo label. The specific process includes: During the pseudo-label update process, the rigid transformation predicted by the model trained in the previous iteration is used to initialize the initial pose of each point cloud pair; The rigid transformation and the corresponding r are calculated using the ICP algorithm. i , if the new r i Less than r stored in the quintuple i , then the quintuple R, t and r i Should be updated to new R, t and r i .

[0019] Preferably, a three-dimensional electronic map of the substation is drawn based on the registered three-dimensional point cloud data by three-dimensional modeling software to achieve three-dimensional modeling of equipment and environment.

[0020] Embodiment 2: This embodiment provides a method for constructing a three-dimensional point cloud map of a substation based on point cloud registration, and its main process is as follows: (1) Field and office work of substation 3D modeling, the flow chart is as follows Figure 2 Field work includes site exploration, laser scanning plan development and data collection, 3D scanning of buildings and equipment to obtain point cloud data. The laser radar scanning diagram is shown in the figure below. Figure 3 shown.

[0021] In the field, data quality needs to be checked in real time, plans need to be adjusted, and data availability needs to be ensured. Due to the large amount of data and redundancy, data preprocessing and 3D point cloud registration are performed by software in the office to optimize data and retain key information.

[0022] (2) Registration calculation of 3D point cloud data. In 3D laser scanning of substations, in order to obtain complete point cloud data, it is necessary to scan from different perspectives and convert the data of each coordinate system into a unified coordinate system. The point cloud registration method is used to calculate the rigid transformation matrix of rotation and translation to achieve the fusion of the source point cloud and the target point cloud.

[0023] (3) Construction of 3D electronic map. The 3D point cloud model of the substation is obtained through field and indoor processing, and the 3D coordinates, color and size data in the model are used for reference modeling. Subsequently, the 3D modeling software is used to draw small parts in proportion to construct a realistic and accurate 3D electronic map of the substation.

[0024] The model was trained and tested on the ModelNet40 dataset. The dataset contains a total of 12,311 CAD models of 40 different categories, of which 9,843 are used in the training set to train the model, and the remaining 2,500 are used in the test set to evaluate the model performance. In the experiment, the performance of the model was verified using the ModelNet40 dataset. To test the registration effect of the model, SEMI-DCP was compared with a variety of registration algorithms (including ICP, Go-ICP, FGR, PointNetLK and DCP). During the training process, the total number of iterations of the model was set to 10. In each iteration, the model network was trained for 250 epochs, and the network parameters were adjusted using the ADAM optimizer with an initial learning rate of 0.001. In order to evaluate the error between the predicted value and the true value, indicators such as mean square error (MSE), root mean square error (RMSE) and mean absolute error (MAE) were used.

[0025] First, the transformer is scanned using a 3D laser scanner. After that, it is processed outdoors and indoors. Finally, a 3D point cloud model of the transformer is generated. The source point cloud is marked in green, and the target point cloud is marked in red. Their initial positions and postures are shown in Figure 2. Figure 4 shown.

[0026] The SEMI-DCP algorithm is applied to the registration of the source point cloud and the target point cloud. The final result is as follows: Figure 5 and Figure 6 shown.

[0027] In the model validity experiment, it is assumed that the training set contains a very small number of labeled samples (300 pairs of point clouds) and 9543 pairs of unlabeled samples. Traditional methods (such as ICP, Go-ICP and FGR) are directly evaluated on the test set containing 2500 pairs of substation clouds. In addition, deep learning models including PointNetLK and DCP are trained on the registered labeled point clouds. The experimental results are shown in Figure 2. Figure 7 shown.

[0028] In the model convergence experiment, Gaussian distributed noise N(0,0.01) is added to the data set to ensure that each point cloud point is interfered by the noise. Figure 8 The experimental results show that ICP and FGR are more sensitive to noise, while Go-ICP and PointNetLK still perform smoothly under the influence of noise. In contrast, the SEMI-DCP model shows excellent noise resistance.

[0029] The preferred embodiments of the present invention are described in detail above in conjunction with the accompanying drawings. However, the present invention is not limited to the specific details in the above embodiments. Within the technical concept of the present invention, the technical solution of the present invention can be subjected to a variety of simple modifications, and these simple modifications all belong to the protection scope of the present invention. It should also be noted that the various specific technical features described in the above specific embodiments can be combined in any suitable manner without contradiction. In order to avoid unnecessary repetition, the present invention will not further describe various possible combinations.

[0030] In addition, various embodiments of the present invention may be arbitrarily combined, and as long as they do not violate the concept of the present invention, they should also be regarded as the contents disclosed by the present invention.

Claims

1. A method for constructing a three-dimensional map of a substation based on point cloud registration, characterized in that: The following steps are involved: S1, using 3D laser scanning equipment to obtain multi-view point cloud data of equipment in the substation, including source point cloud and target point cloud ,get N The point cloud collection is ; The registered point cloud data is , the unregistered point cloud data is ; S2, using the iterative closest point algorithm ICP to generate pseudo labels for unregistered point cloud data; S3, selects the generated pseudo labels and combines them with the real label point cloud data, and uses a semi-supervised deep neural network point cloud registration model to train and optimize the point cloud data; S4, through the iterative optimization of the semi-supervised deep neural network point cloud registration model DCP, gradually improves the accuracy of pseudo-labels.

2. A method for constructing a three-dimensional map of a substation based on point cloud registration according to claim 1, characterized in that: In step S1, a three-dimensional laser scanning device is used to obtain multi-view point cloud data of equipment in the substation, including the following process: Preliminary preparations were made by surveying the site, developing laser scanning plans and data collection procedures, and conducting 3D laser scanning of the buildings and equipment in the station to obtain 3D point cloud data; Set up multiple points to perform all-round scanning of the target from different perspectives and positions, perform registration calculations on the point cloud data, and convert the 3D point cloud data in the independent coordinate systems of different points into a unified coordinate system; The three-dimensional point cloud model of the substation obtained through field and indoor processing is modeled in combination with its three-dimensional coordinates, color and size data. The small components are drawn in proportion using three-dimensional modeling software to construct a real and accurate three-dimensional electronic map of the substation.

3. The method for constructing a three-dimensional map of a substation based on point cloud registration according to claim 2, characterized in that: In step S2, the iterative closest point algorithm is used to generate pseudo labels for the unregistered point cloud data, including: Input unregistered point cloud dataset For the point cloud pairs in , set the threshold tau for algorithm termination; Calculate the closest point, for the i Point Cloud , under the condition of minimizing the neighboring distance, the following equation is solved and : ； in, is the rigid transformation matrix, R represents the rotation matrix predicted by the model, t represents the predicted translation vector; Calculate the rigid transformation. If the optimal correspondence between the source point cloud and the target point cloud is known, the optimal rigid transformation can be calculated as follows: ； in, To minimize the mean square error, and ; Target point cloud Medium distance The nearest point; Update the current point cloud and calculate the error between the current point cloud and the adjacent point cloud; determine whether the error is less than the threshold tau, if so, terminate the iteration, otherwise continue the iterative optimization calculation.

4. The method for constructing a three-dimensional map of a substation based on point cloud registration according to claim 3 is characterized in that: In step S3, when training the deep neural network registration model, from the set without labels A subset of samples with pseudo labels is selected from the dataset to increase the number of labeled samples to reduce overfitting during model training. The selection of generated pseudo labels includes the following process: Train the DCP model on the registered point cloud data to obtain the initial model parameters ,parameter As the pre-trained parameters of the model SEMI-DCP, Indicates The values ​​of the model parameters at the iteration; Calculate the root mean square error when registering unlabeled point clouds ; Root mean square error For each point cloud pair, a tuple of five elements is formed ; according to Sort all tuples in ascending order of value; Select the first k pairs of point clouds and the corresponding pseudo labels together with the registered point cloud data Participate in the training of SEMI-DCP model; Finally, k pairs of pseudo-label samples are output .

5. The method for constructing a three-dimensional map of a substation based on point cloud registration according to claim 4, characterized in that: The specific process of step S4 includes: During the pseudo-label update process, the rigid transformation predicted by the model trained in the previous iteration is used to initialize the initial pose of each point cloud pair; The rigid transformation and the corresponding r are calculated using the ICP algorithm. i , if the new r i Less than r stored in the quintuple i , then the quintuple R, t and r i Should be updated to new R, t and r i .

6. A method for constructing a three-dimensional map of a substation based on point cloud registration according to claim 5, characterized in that: The 3D modeling software is used to draw the 3D electronic map of the substation based on the registered 3D point cloud data to achieve 3D modeling of equipment and environment.

Images