Safety intelligent monitoring method for transformer substation portal frame drainage wire

Abstract

The invention relates to the technical field of power line monitoring, and discloses a transformer substation portal frame drainage wire safety intelligent monitoring method. Comprising the steps of obtaining accurate point cloud information of the portal frame drainage wire, preprocessing the collected point cloud data, projecting the spherical surface of the portal frame drainage wire to a plane to obtain a two-dimensional image, training by adopting an improved lightweight deep learning network, accurately segmenting pixel points of the drainage wire from the point cloud data, and performing feature fitting on the drainage wire. And calculating the distance between the two drainage wires. Compared with a method of directly using three-dimensional point cloud data for segmentation, the method provided by the invention has the advantages that the data volume is reduced, the labeling difficulty is reduced, and the prediction speed and precision are balanced, so that the function of performing real-time early warning on edge equipment deployment can be realized. According to the method, the distance between portal frame drainage wires is detected, potential safety hazards are found in time, early warning signals are triggered, and safe operation of a transformer substation is ensured.

Description

Technical Field

[0001] The invention relates to the technical field of power line monitoring, and in particular to a method for intelligently monitoring the safety of a gantry drainage line of a transformer substation. Background Art

[0002] The increase in electricity demand and the expansion of the scale of power systems have put forward higher requirements for the safety and stability of power systems. In power systems, transformers are one of the important equipment in power systems and undertake the important task of voltage conversion. The safety and stability of their operation directly affect the stability of the entire power system. There is a certain margin between the gantry drainage lines, but under complex environmental conditions, the two-phase drainage lines will deviate in a short period of time, which will cause inter-phase short-circuit faults due to the action of short-circuit electromotive force, causing the main transformer to trip, destroying the stability of the power system, and endangering the safe and stable operation of the power grid.

[0003] However, traditional monitoring and maintenance methods often rely on manual inspections, which are inefficient, costly, prone to missed detections and misjudgments, and cannot be guaranteed in terms of immediacy. The use of laser point cloud technology for real-time intelligent safety detection is of great significance to maintaining the safety of substation systems.

[0004] In the power line monitoring and early warning system, an existing technical solution is implemented based on point cloud data processing and deep learning models. This method obtains high-quality point cloud data by removing outliers, thinning and registering the data set. Then, based on the line detection and region growing segmentation algorithm, the point cloud data is segmented, the power line point cloud is extracted, and finally the power line point cloud is classified.

[0005] Existing power line fitting techniques are mainly divided into three categories: original point cloud-based, voxel-based, and projection-based methods. The original point cloud-based method directly uses the original point cloud as input; however, this method has high requirements for computational efficiency and memory, and it is difficult to process large-scale point cloud data. Point cloud voxelization refers to segmenting irregular point clouds into regular voxels, and then using three-dimensional convolution for feature learning; although this method can clearly encode three-dimensional spatial information, since the input is a fixed grid of voxels, all points in the same voxel have the same semantic label; in addition, as the grid resolution increases, the memory and computational complexity increase exponentially, and fine-grained reasoning cannot be achieved in large-scale scenes. The projection-based method projects the point cloud onto a plane or curved surface to generate a dense and compact two-dimensional projection map; although this method can flexibly adapt to the scale of different scenes, in order to improve the segmentation accuracy, the existing technology usually achieves it by increasing the complexity of the model, and it is difficult to strike a balance between segmentation accuracy and speed. In addition, the existing technology is limited to fitting power lines, and fails to achieve further early warning functions, and there is a gap in the safety early warning of the gantry drainage line of the substation. Summary of the invention

[0006] In order to solve the above technical deficiencies and meet the requirements for algorithm accuracy and real-time performance in substation scenarios; the present invention proposes a safe and intelligent monitoring method for substation gantry drainage lines. The method is based on spherical projection of laser point cloud, spherically projects laser point cloud data onto a two-dimensional plane, and uses a real-time semantic segmentation network to balance detection speed, detection accuracy, equipment computing power and storage requirements, thereby improving the segmentation accuracy of gantry drainage line extraction; in addition, a warning module is used to issue a real-time warning of drainage line short-circuit conditions to ensure the normal operation of the substation.

[0007] The technical solution adopted by the present invention is:

[0008] A method for intelligently monitoring the safety of a gantry drainage line of a substation, the method for intelligently monitoring the safety of a gantry drainage line of a substation is based on spherical projection of a laser point cloud, and comprises the following steps:

[0009] Step 1. Use laser radar equipment to collect three-dimensional point cloud data at the substation site to obtain accurate point cloud information of the gantry drainage line;

[0010] Step 2. Preprocess the collected point cloud data, including point cloud denoising, filtering and ground filtering, to improve the quality and usability of the data;

[0011] Step 3. After filtering out the ground, the gantry drainage line is spherically projected onto the plane to obtain a two-dimensional image, which is manually annotated and a gantry drainage line dataset is constructed;

[0012] Step 4. Based on the constructed gantry drainage line dataset, an improved lightweight deep learning network is used for training to enhance the accuracy and efficiency of the model;

[0013] Step 5. Apply the trained deep learning model to accurately segment the pixels of the drainage lines from the point cloud data and post-process the segmentation results;

[0014] Step 6. Based on the segmentation results of the deep learning model, feature fitting is performed on the drainage lines and the distance between the two drainage lines is calculated; once the distance exceeds the preset threshold, the system will issue an alarm signal in real time.

[0015] Further, step 1 includes the following steps:

[0016] Step 1.1, installation and parameter setting of LiDAR equipment:

[0017] Install the LiDAR equipment at key locations in the substation to ensure that the field of view can fully cover the gantry drainage line; set the scanning parameters of the LiDAR, including scanning frequency, angle range and resolution, to ensure high-quality and efficient collection of point cloud data;

[0018] Step 1.2, point cloud data collection and environmental parameter recording:

[0019] For monitoring scenarios, multiple sets of point cloud data are collected, and relevant parameters and environmental information during the collection process are recorded in detail.

[0020] Furthermore, in step 2, the point cloud denoising process is as follows:

[0021] Gaussian distribution filtering is used for denoising. By analyzing the neighborhood of each point, the mean and standard deviation of the Gaussian distribution are used to determine the threshold. Points exceeding the threshold are considered as noise and removed.

[0022] The one-dimensional form of the Gaussian function is:

[0023]

[0024] In the formula, f(x) represents the probability density value at position x; σ is the function smoothness parameter; Normalization factor; e base of the natural logarithm; is the exponential part of the Gaussian function; μ represents the expected value of the normal distribution;

[0025] Convert the value x to a two-dimensional vector (x, y). The two-dimensional form of the Gaussian filter is:

[0026]

[0027] Where f(x,y) represents the weight of the probability density value in two-dimensional space; 2πσ 2 is the standardization factor; σ is the function smoothness parameter; e is the base of the natural logarithm; is the exponential part; x 2 +y 2 is the square of the Euclidean distance from a point to the origin;

[0028] The calculation process is as follows:

[0029] Calculate the distance from the point cloud in the neighborhood to the center point, and substitute it into the Gaussian function to generate a Gaussian template; if the Gaussian template contains a small value, normalize the generated Gaussian template so that the value of its upper left corner is 1; align the center of the Gaussian template with the point cloud matrix to be processed; multiply and add the corresponding elements, and fill the positions with zeros where there are no elements;

[0030] The above calculation is performed on each point cloud element to finally generate a denoised output matrix.

[0031] Further, in step 2, the ground filtering process is as follows:

[0032] The RANSAC method is used to filter the ground. After obtaining the shape of the desired object in advance, the parameters representing the mathematical model are solved iteratively to finally obtain a point cloud set representing the object.

[0033] The ground is a plane in space, and its equation is expressed as follows:

[0034] Axx+Bxy+Cxz+D=0

[0035] In the formula, A is the inclination coefficient of the plane in the x direction; B is the inclination coefficient in the y direction; C is the inclination coefficient of the plane in the z direction; D is a constant term; x, y, z are coordinate axis variables in three-dimensional space; A, B, C are not 0 at the same time;

[0036] Randomly select three points p in the substation cloud 1 (x 1 ,y 1 ,z 1 ), p 2 (x 2 ,y 2 ,z 2 ), p 3 (x 3 ,y 3 ,z 3 ), solve the plane parameter value according to the following formula:

[0037]

[0038] In the formula, A, B, C are the coefficients of the plane equation; D is the constant term;

[0039] Calculate the plane distance d i , the calculation is shown as follows:

[0040]

[0041] Where, d i Point p i (x i ,y i ,z i ) to the plane; A, B, C are the coefficients of the plane equation; xi, yi, zi are the coordinates of pi in three-dimensional space; D is a constant term;

[0042] According to the obtained plane model, points whose plane distance d is greater than the threshold are outliers, and points whose distance is less than the threshold are inliers; the threshold is usually 0.01 to 0.1;

[0043] Update the parameter model according to the number of points in the local area, and continue to iterate until the specified number of times is reached;

[0044] The number of iterations k is calculated by the following formula:

[0045]

[0046] Where z is the ratio of the in-place points to the total number of point clouds; p is the probability that the sample points selected during the iteration are in-place points; n is the number of points required to simulate the plane during the iteration;

[0047] After the iteration is completed, the plane with the most points in the local area is taken as the ground, and the ground points are filtered out from the point cloud data.

[0048] Furthermore, in step 3, the unstructured three-dimensional point cloud data is mapped into a two-dimensional depth image for use in training a convolutional neural network, including the following steps:

[0049] Step 3.1, map the 3D image to 2D:

[0050] Each original point cloud (x, y, z) is mapped to a two-dimensional coordinate (u, v) by the following formula:

[0051]

[0052] Where w and h represent the width and height of the projected image; r represents the distance between each point and the radar; f down is the vertical field of view below the LiDAR sensor; f up is the vertical field of view above the LiDAR sensor; f is the overall vertical field of view of the sensor;

[0053] The 3D point cloud coordinates (x, y, z), intensity values ​​(i) and distance values ​​(r) are stored in the RV image channel respectively, and the final generated [w×h×5] projection image is input into the neural network for learning;

[0054] Step 3.2: Build the dataset:

[0055] The annotation system is used to accurately annotate the gantry drainage lines in the two-dimensional image at the pixel level; and a point cloud segmentation dataset is constructed with 70% of the data used for training, 20% of the data used for verification, and 10% of the data used for testing.

[0056] Furthermore, in step 4, the lightweight deep learning network is a lightweight semantic segmentation model; the lightweight semantic segmentation model uses MobileNetV3 as a feature extraction network; the lightweight semantic segmentation model integrates global information through a context module, uses a pixel shuffle layer Pixel-Shuffle for upsampling, and combines the feature maps of the encoder and decoder in the decoding stage through a jump connection; after the jump connection, the deep and shallow information are further fused;

[0057] The lightweight semantic segmentation model introduces a proxy attention AA module, which expands the traditional attention mechanism's quadruple Q, K, V, A by introducing an additional proxy vector A. In the new mechanism, the proxy vector A acts as a proxy for the query vector Q, aggregating information from the key vector K and the value vector V, and feeding the aggregated information back to the query vector Q.

[0058] Furthermore, based on the lightweight semantic segmentation model, the calculation formula of the upsampling method is as follows:

[0059] PS(T) x,y,c =T [x / r],[y / r],C×mod(x,r)+c

[0060] Where PS is the periodic pixel shuffle operator; T is the output high-resolution feature map; x, y are the positions of pixels in the output high-resolution feature map; c is the number of channels of the input feature map; T [x / r],[y / r] Indicates that the coordinates x and y of the output feature map are mapped back to the coordinates of the input feature map through integer division operation; C×mod(x,r) performs a modulus operation between x and r to obtain the corresponding channel position;

[0061] After PS operation, the input is [H×W×r 2 C] will become [rH×rW×C], where r is the upsampling factor;

[0062] During the training process, an optimized high-resolution feature map can be obtained by learning the weights of r×r Shuffle channels.

[0063] Furthermore, based on the lightweight semantic segmentation model, the model training process is as follows:

[0064] Model training is performed based on the point cloud segmentation dataset constructed in step 3;

[0065] The moment estimation optimizer is used as the training optimization strategy for the overall branch. During the training process, a total of 100 rounds are performed. The initial learning rate is set to 0.0003, and the learning rate is dynamically adjusted after each round of update, multiplied by 0.9. The model loss is monitored, and the training is stopped early when the loss value does not decrease for 5 consecutive rounds.

[0066] Through the above training method, multiple experiments are carried out to obtain multiple network weight parameter files; the model segmentation accuracy of different weight parameter files is compared on the test set, and the weight parameters with the highest accuracy are selected as the network parameters of the final model.

[0067] Furthermore, in step 5, during the data post-processing, the GPU-based K nearest neighbor (KNN) algorithm is introduced for rapid correction. The KNN algorithm searches for the nearest K points in the Euclidean space and then performs a consensus vote to correct the label of the point. Finally, the segmented pixel points are fitted into drainage lines using a polynomial fitting method.

[0068] Furthermore, in step 6, the distance between drainage lines is calculated as follows:

[0069]

[0070] In the formula, D(A,B) represents the Hausdorf f distance between drainage line set A and drainage line set B; A is the point set of the first drainage line; B is the point set of the second drainage line; a represents any point in set A; b represents any point in set B; d(ab) represents the distance between point a and point b;

[0071] The threshold is set according to the actual situation. If the Hausdorff distance is less than the threshold, an alarm is triggered to notify relevant personnel to take necessary safety measures.

[0072] Compared with the existing technology, the substation gantry drainage line safety intelligent monitoring method has the following beneficial effects:

[0073] The intelligent monitoring method for the safety of the gantry drainage line of the substation is based on the spherical projection of the laser point cloud. By spherically projecting the three-dimensional point cloud data onto a two-dimensional plane, compared with the method of directly using the three-dimensional point cloud data for segmentation, it reduces the amount of data, reduces the difficulty of annotation, and balances the prediction speed and accuracy, so as to realize the function of real-time early warning on edge device deployment. Compared with the plane projection, the spherical projection can better preserve the spatial relationship in the point cloud data and improve the segmentation accuracy.

[0074] The intelligent monitoring method for the safety of the gantry drainage line of the substation adopts a lightweight semantic segmentation model, an improved backbone structure and a proxy attention mechanism, which improves the model's expressiveness without increasing the computational complexity. By introducing additional proxy vectors, the proxy attention mechanism achieves efficient global information modeling and significantly improves the accuracy of small target detection.

[0075] The intelligent safety monitoring method for the gantry drainage line of the substation uses a pixel shuffle layer to replace the traditional transposed convolution, realizing an efficient and lightweight decoding process, avoiding the checkerboard effect, and improving the decoding efficiency and segmentation accuracy.

[0076] In summary, the intelligent safety monitoring method for the gantry drainage line of the substation can run efficiently on the edge device. By detecting the distance between the gantry drainage lines, safety hazards can be discovered in time and early warning signals can be triggered to ensure the safe operation of the substation. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative labor.

[0078] in:

[0079] Figure 1 It is a flow chart of the method for intelligent safety monitoring of the gantry drainage line of a substation according to the present invention;

[0080] Figure 2 It is a judgment logic diagram of the substation gantry drainage line safety intelligent monitoring method of the present invention. DETAILED DESCRIPTION

[0081] 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.

[0082] The present invention proposes a method for intelligently monitoring the safety of a gantry drainage line of a substation; the process of the method for intelligently monitoring the safety of a gantry drainage line of a substation is as follows: Figure 1 As shown, the judgment logic is as follows Figure 2 As shown, the specific steps include:

[0083] Step 1. Use laser radar equipment to collect three-dimensional point cloud data at the substation site to obtain accurate point cloud information of the gantry drainage line; specifically, the following steps are included:

[0084] Step 1.1, installation and parameter setting of LiDAR equipment:

[0085] Install the LiDAR equipment at key locations in the substation to ensure that the field of view can fully cover the gantry drainage line. Set the scanning parameters of the LiDAR, including scanning frequency, angle range, and resolution, to ensure high quality and efficient collection of point cloud data.

[0086] Step 1.2, point cloud data collection and environmental parameter recording:

[0087] For monitoring scenarios, multiple sets of point cloud data are collected, and relevant parameters and environmental information during the collection process are recorded in detail to provide a reference for subsequent data processing and analysis.

[0088] Step 2. Preprocess the collected point cloud data, including point cloud denoising, filtering and ground filtering, to improve the quality and usability of the data; the specific process is as follows:

[0089] Step 2.1, point cloud denoising:

[0090] Due to the different roughness of the scanned object surface and the influence of noise in the external environment, the point cloud data may contain some small point clouds and discrete points that are different from the main point cloud features and far away from the main point cloud. These noises are not conducive to the extraction and matching of point cloud features.

[0091] In order to solve this problem, this embodiment uses a Gaussian distribution filtering method to denoise the original point cloud data to improve the accuracy of subsequent registration of the laser point cloud and the visible light image.

[0092] Specifically, the Gaussian distribution filtering method performs statistical analysis on the neighborhood of each point, uses the mean and standard deviation of the Gaussian distribution to determine the threshold, and treats points exceeding the threshold as noise points for elimination.

[0093] The one-dimensional form of the Gaussian function is:

[0094]

[0095] In the formula, f(x) represents the probability density value at position x; σ is the function smoothness parameter, the larger the σ, the better the smoothness; Normalization factor, used to ensure that the area of ​​the probability density function is equal to 1; e is the base of the natural logarithm, approximately equal to 2.71828; It is the exponential part of the Gaussian function, which represents the ratio of the square of the distance from the mean μ to the square of the standard deviation, reflecting the distance of x from the mean; μ represents the expected value of the normal distribution and determines the symmetry center of the distribution function.

[0096] Convert the value x to a two-dimensional vector (x, y). The two-dimensional form of the Gaussian filter is:

[0097]

[0098] Where f(x,y) represents the weight of the probability density value in two-dimensional space; 2πσ 2 is a normalization factor used to ensure that the integral of the entire two-dimensional Gaussian function is equal to 1; σ is the smoothness parameter of the function; e is the base of the natural logarithm, which is approximately equal to 2.71828; is the exponential part, which represents the ratio of the square distance from the origin (0,0) to the square of the standard deviation; x2 +y 2 It is the square of the Euclidean distance from a point to the origin, reflecting how far the point is from the center point on the plane.

[0099] The above two-dimensional Gaussian function is used to perform linear convolution on the neighborhood of each point cloud. Point clouds at different positions in the neighborhood are assigned different weights. The closer the point is, the greater the weight, and the farther the point is, the smaller the weight.

[0100] The specific calculation process is as follows:

[0101] Calculate the distance from other point clouds in the neighborhood to the center of the neighborhood, and substitute the distance into the Gaussian function to generate a Gaussian template;

[0102] If the Gaussian template contains small values, it is normalized and the value of the upper left corner of the template is normalized to 1;

[0103] Align the center of the Gaussian template with the point cloud matrix to be processed, and perform a linear convolution operation; that is, multiply and add the corresponding elements, and fill zeros for positions where there are no elements;

[0104] The above calculation is performed on each point cloud element to finally generate a denoised output matrix.

[0105] Step 2.2, ground filtering:

[0106] This embodiment uses the RANSAC method to filter out the ground; RANSAC is an algorithm based on statistical probability. After obtaining the shape of the desired object in advance, it solves the parameters representing the mathematical model in an iterative manner to finally obtain a point cloud set representing the object.

[0107] The ground is a plane in space, and its equation is expressed as follows:

[0108] Axx+Bxy+Cxz+D=0

[0109] In the formula, A is the inclination coefficient of the plane in the x direction; B is the inclination coefficient in the y direction; C is the inclination coefficient of the plane in the z direction; D is a constant term used to adjust the relative position of the plane and the origin; x, y, z are coordinate axis variables in three-dimensional space, representing the coordinate values ​​in the x, y and z directions; where A, B, C are not 0 at the same time.

[0110] Randomly select three points p in the substation cloud 1 (x 1 ,y 1 ,z 1 ), p 2 (x 2 ,y 2 ,z 2 ), p 3 (x3 ,y 3 ,z 3 ), solve the plane parameter value according to the following formula:

[0111]

[0112] In the formula, A, B, and C are the coefficients of the plane equation, representing the components of the plane's normal vector; D is a constant term that controls the plane's height in space and affects the distance between the plane and the origin.

[0113] According to the obtained plane model, points whose plane distance d is greater than the threshold are outliers, and points whose distance is less than the threshold are inliers; the threshold is usually 0.01 to 0.1.

[0114]

[0115] Where, d i Point p i (x i ,y i ,z i ) is the distance from the plane; A, B, C are the coefficients of the plane equation; x i ,y i ,z i The point is p i Coordinates in three-dimensional space; D is a constant term; |A×x i +B×y i +C×z i +D| is the calculation point p i (x i ,y i ,z i ) to the plane; Calculate the magnitude of the plane normal vector;

[0116] Each iteration calculates the error between the local points and the parameter model, then updates the data and calculates the parameter model again until the iteration is completed.

[0117] The number of iterations k is calculated by the following formula:

[0118]

[0119] Where z is the ratio of in-library points to the total number of point clouds; p is the probability that the sample points selected during the iteration are in-library points; and n is the number of points required to simulate the plane during the iteration.

[0120] After the iteration is completed, the plane with the most points in the local area is taken as the ground, and the ground points are filtered out from the point cloud data.

[0121] Step 3. After filtering out the ground, the gantry drainage line is spherically projected onto the plane to obtain a two-dimensional image, which is manually annotated and a gantry drainage line dataset is constructed. The specific process is as follows:

[0122] Step 3.1, map the 3D image to 2D:

[0123] The unstructured 3D point cloud is spherically projected to obtain the original depth image of the lidar, which can be input into a standard convolutional neural network for learning.

[0124] In the two-dimensional projection image RV, each original point cloud (x, y, z) is mapped to a two-dimensional coordinate (u, v) by the following formula:

[0125]

[0126] In the formula, w and h represent the width and height of the projected image; r represents the distance between each point and the radar. The calculation formula is: f down is the vertical field of view below the LiDAR sensor; f up is the vertical field of view above the lidar sensor; f is the overall vertical field of view of the sensor, defined as: f = |f down |+|f up |.

[0127] In order to retain the complete field of view and spatial information of the rotating lidar as much as possible, the 3D point cloud coordinates (x, y, z), intensity values ​​(i) and distance values ​​(r) are stored in the RV image channel respectively, and finally the [w×h×5] projection image is generated and input into the neural network for learning.

[0128] Step 3.2: Build the dataset:

[0129] The annotation system is used to accurately annotate the gantry drainage lines in the two-dimensional image at the pixel level; and a point cloud segmentation dataset is constructed with 70% of the data used for training, 20% of the data used for verification, and 10% of the data used for testing.

[0130] Through the above steps, the idea of ​​spherically projecting the laser point cloud onto a plane to segment the gantry drainage line is realized. The redundant three-dimensional data is projected onto a two-dimensional plane, and the segmentation model is input for training to reduce the amount of point cloud data, meet the computing power of edge devices and the real-time requirements of substation scenarios, and ensure segmentation accuracy. The laser point cloud data is spherically projected onto a two-dimensional plane to reduce data complexity and meet real-time requirements. This method not only effectively reduces the amount of data input to the model, but also ensures the accuracy and real-time requirements of the segmentation model in the substation scenario.

[0131] Step 4. Based on the constructed gantry drainage line dataset, an improved lightweight deep learning network is used for training to enhance the accuracy and efficiency of the model; the specific process is as follows:

[0132] Step 4.1, design a lightweight semantic segmentation model:

[0133] The lightweight semantic segmentation network mentioned in step 4 follows the hourglass-shaped architecture of the encoder and decoder. The global information is aggregated by the context module of the head network. The encoder uses the lightweight network MobilenetV3 for feature extraction. The decoder module uses a pixel shuffle layer for upsampling and fused the features extracted in the downsampling process with the feature map of the same size of the encoder through jumps. In order to further utilize the spatial detail information, the deep and shallow information is fused through the feature fusion module after the jump connection.

[0134] In the substation scene, there is a serious imbalance between foreground objects such as gantry drain wires and background objects such as substation facilities and buildings, which makes the traditional semantic segmentation network face challenges in the feature extraction stage. Especially when dealing with small objects, due to their relatively small number of pixels, it is easy to lose key information during the downsampling process.

[0135] To address this problem, this embodiment introduces an Agent Attention (AA) module. The underlying features of the image contain rich contour, edge and position information, which is crucial for accurately identifying the gantry drainage line.

[0136] However, traditional convolution operations may not be able to adequately capture sufficient contextual information when processing these features, so a proxy attention mechanism is adopted to enhance the feature extraction process.

[0137] The AgentAttention (AA) module extends the traditional attention mechanism's quadruple (Q, K, V, A) by introducing an additional agent vector A. In this new mechanism, the agent vector A acts as a proxy for the query vector Q, aggregating information from the key vector K and the value vector V, and feeding the aggregated information back to the query vector Q.

[0138] This module can not only effectively capture the key features of the gantry drainage line, but also improve the processing speed while maintaining computational efficiency. The proxy attention mechanism combines the advantages of the Softmax attention mechanism and the linear attention mechanism, with low computational complexity and high model expression ability; its characteristics enable the model to extract features more accurately and maintain efficient computing speed when processing small foreground targets in substation scenarios.

[0139] Semantic segmentation networks based on codec structures usually use transposed convolution to achieve upsampling of feature maps. However, transposed convolution not only requires a large amount of video memory during training, but also causes a checkerboard effect due to uneven overlap during the mapping process. This embodiment uses a pixel shuffle layer instead of a transposed convolution for upsampling.

[0140] The Pixel-Shuffle layer generates a high-resolution feature map by randomly arranging the learned feature maps in the channel dimension to the spatial dimension. This process is also called the sub-pixel convolution process. Specifically, Pixel-Shuffle treats each pixel of the low-resolution image as multiple sub-pixels, and interpolates the sub-pixels in multiple channels according to certain rules to achieve pixel reorganization. Sub-pixel convolution periodically activates the pixel shuffle operation according to the positions of different sub-pixels.

[0141] The calculation formula is as follows:

[0142] PS(T) x,y,c =T [x / r],[y / r],C×mod(x,r)+c

[0143] Where PS is the periodic pixel shuffle operator; T is the output high-resolution feature map; x, y are the positions of pixels in the output high-resolution feature map; c is the number of channels of the input feature map; T [x / r],[y / r] It means that the coordinates x and y of the output feature map are mapped back to the coordinates of the input feature map through integer division operation; C×mod(x,r) performs a modulus operation between x and r to obtain the corresponding channel position.

[0144] After PS operation, the input is [H×W×r 2 C] will become [rH×rW×C], where r is the upsampling factor;

[0145] During the training process, an optimized high-resolution feature map can be obtained by learning the weights of r×r Shuffle channels.

[0146] Step 4.2, model training:

[0147] The laser point cloud gantry drainage line detection model needs to be trained before segmentation; the steps are as follows:

[0148] a) Based on the point cloud segmentation dataset constructed in step 3, the training was performed using the moment estimation optimizer as the training optimization strategy for the overall branch, with a total of 100 rounds of training. The learning rate was dynamically adjusted starting from 0.0003, and the learning rate was multiplied by 0.9 after each round of update. The model loss was calculated after each round of training, and the training was stopped when the loss value did not decrease for 5 consecutive rounds.

[0149] b) Use the training method in step a) to train the model multiple times, obtain multiple network weight parameter files of the model through multiple training, compare the model segmentation accuracy using different weight parameter files under the test set, and select the network weight parameters with the highest accuracy as the parameters of each network in the model.

[0150] In step 4, a lightweight semantic segmentation model was designed, using a lightweight backbone structure suitable for edge devices and introducing an attention mechanism that balances accuracy and speed. This design not only improves the computational efficiency of the model in the inference phase, but also maintains a high level of segmentation accuracy, thereby achieving real-time detection of the gantry drainage line.

[0151] Step 5. Apply the trained deep learning model to accurately segment the pixels of the drainage lines from the point cloud data and post-process the segmentation results. The specific process is as follows:

[0152] Step 5.1, data post-processing:

[0153] After obtaining the pixel-by-pixel classification results of the projection image, the segmentation results need to be mapped back to the three-dimensional point cloud to obtain point-by-point labels. However, due to the segmentation error of the projection image and the fact that one pixel label often corresponds to multiple point clouds during the label mapping process, classification errors at the boundaries and ambiguous semantic segmentation results may occur. To solve this problem, the present invention introduces a GPU-based K-nearest neighbor (KNN) algorithm for rapid correction.

[0154] The KNN algorithm searches for the nearest K points in Euclidean space and then performs a consensus vote to correct the label of the point.

[0155] Step 5.2: Drainage line fitting:

[0156] The segmented pixel points are fitted into drainage lines using a polynomial fitting method.

[0157] Step 6. Based on the segmentation results of the deep learning model, feature fitting is performed on the drainage lines and the distance between the two drainage lines is calculated; once the distance exceeds the preset threshold, the system will issue an alarm signal in real time.

[0158] Step 6.1, calculate the distance between drainage lines:

[0159] Calculate the distance between the two drain wires and calculate the Hausdorff distance between the two wires, that is, the minimum value of the maximum distance from each drain wire to the other drain wire. The calculation formula is as follows:

[0160]

[0161] Where D(A,B) represents the Hausdorff distance between drainage line set A and drainage line set B; A is the point set of the first drainage line; B is the point set of the second drainage line; a represents any point in set A; b represents any point in set B; d(ab) represents the distance between point a and point b.

[0162] Step 6.2, judgment and alarm:

[0163] The threshold is set according to the actual situation. If the Hausdorff distance is less than the threshold, an alarm is triggered to notify relevant personnel to take necessary safety measures.

[0164] In summary, the safety intelligent monitoring method for the gantry drainage line of the substation, combined with point cloud data and deep learning models, has formed a set of safety intelligent monitoring methods specifically used for real-time detection and monitoring of the gantry drainage lines of the substation. It can timely detect the abnormal distance between the gantry drainage lines, thereby preventing potential safety risks and ensuring the safe and continuous operation of the substation.

[0165] It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above and that the invention can be implemented in other specific forms without departing from the spirit or essential features of the invention. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description, and it is intended that all variations falling within the meaning and scope of the same elements of the claims be included in the invention. Any reference numeral in a claim should not be considered as limiting the claim to which it relates.

Claims

1. A method for intelligent safety monitoring of a substation gantry drainage line, characterized in that: The substation gantry drainage line safety intelligent monitoring method is based on laser point cloud spherical projection, and includes the following steps: Step 1. Use laser radar equipment to collect three-dimensional point cloud data at the substation site to obtain accurate point cloud information of the gantry drainage line; Step 2. Preprocess the collected point cloud data, including point cloud denoising, filtering and ground filtering, to improve the quality and usability of the data; Step 3. After filtering out the ground, the gantry drainage line is spherically projected onto a plane to obtain a two-dimensional image, which is manually annotated and a gantry drainage line dataset is constructed; Step 4. Based on the constructed gantry drainage line dataset, an improved lightweight deep learning network is used for training to enhance the accuracy and efficiency of the model; Step 5. Apply the trained deep learning model to accurately segment the pixels of the drainage lines from the point cloud data and post-process the segmentation results; Step 6. Based on the segmentation results of the deep learning model, feature fitting is performed on the drainage lines and the distance between the two drainage lines is calculated; once the distance exceeds the preset threshold, the system will issue an alarm signal in real time.

2. The method for intelligent safety monitoring of the gantry drainage line of a substation according to claim 1 is characterized in that: Step 1 includes the following steps: Step 1.1, installation and parameter setting of LiDAR equipment: Install the LiDAR equipment at key locations in the substation to ensure that the field of view can fully cover the gantry drainage line; set the scanning parameters of the LiDAR, including scanning frequency, angle range and resolution, to ensure high-quality and efficient collection of point cloud data; Step 1.2, point cloud data collection and environmental parameter recording: For monitoring scenarios, multiple sets of point cloud data are collected, and relevant parameters and environmental information during the collection process are recorded in detail.

3. The method for intelligent safety monitoring of the gantry drainage line of a substation according to claim 1 is characterized in that: In step 2, the point cloud denoising process is as follows: Gaussian distribution filtering is used for denoising. By analyzing the neighborhood of each point, the mean and standard deviation of the Gaussian distribution are used to determine the threshold. Points exceeding the threshold are considered as noise and removed. The one-dimensional form of the Gaussian function is: In the formula, f(x) represents the probability density value at position x; σ is the function smoothness parameter; Normalization factor; e base of the natural logarithm; is the exponential part of the Gaussian function; μ represents the expected value of the normal distribution; Convert the value x to a two-dimensional vector (x, y). The two-dimensional form of the Gaussian filter is: Where f(x,y) represents the weight of the probability density value in two-dimensional space; 2πσ 2 is the standardization factor; σ is the function smoothness parameter; e is the base of the natural logarithm; is the exponential part; x 2 +y 2 is the square of the Euclidean distance from a point to the origin; The calculation process is as follows: Calculate the distance from the point cloud in the neighborhood to the center point, and substitute it into the Gaussian function to generate a Gaussian template; if the Gaussian template contains a small value, normalize the generated Gaussian template so that the value of its upper left corner is 1; align the center of the Gaussian template with the point cloud matrix to be processed; multiply and add the corresponding elements, and fill the positions with zeros where there are no elements; The above calculation is performed on each point cloud element to finally generate a denoised output matrix.

4. The method for intelligent safety monitoring of the substation gantry drainage line according to claim 1 is characterized in that: In step 2, the ground filtering process is as follows: The RANSAC method is used to filter the ground. After obtaining the shape of the desired object in advance, the parameters representing the mathematical model are solved iteratively to finally obtain a point cloud set representing the object. The ground is a plane in space, and its equation is expressed as follows: Axx+Bxy+Cxz+D=0 In the formula, A is the inclination coefficient of the plane in the x direction; B is the inclination coefficient in the y direction; C is the inclination coefficient of the plane in the z direction; D is a constant term; x, y, z are coordinate axis variables in three-dimensional space; A, B, C are not 0 at the same time; Randomly select three points p1(x1,y1,z1), p2(x2,y2,z2), and p3(x3,y3,z3) in the substation cloud, and solve the plane parameter values ​​according to the following formula: In the formula, A, B, C are the coefficients of the plane equation; D is the constant term; Calculate the plane distance d i , the calculation is shown as follows: Where, d i Point p i (x i ,y i ,z i ) is the distance from the plane; A, B, C are the coefficients of the plane equation; x i ,y i ,z i The point is p i Coordinates in three-dimensional space; D is a constant term; According to the obtained plane model, points whose plane distance d is greater than the threshold are outliers, and points whose distance is less than the threshold are inliers; the threshold is usually 0.01 to 0.1; Update the parameter model according to the number of points in the local area, and continue to iterate until the specified number of times is reached; The number of iterations k is calculated by the following formula: Where z is the ratio of the in-place points to the total number of point clouds; p is the probability that the sample points selected during the iteration are in-place points; n is the number of points required to simulate the plane during the iteration; After the iteration is completed, the plane with the most points in the local area is taken as the ground, and the ground points are filtered out from the point cloud data.

5. The method for intelligent safety monitoring of the gantry drainage line of a substation according to claim 1 is characterized in that: In step 3, the unstructured three-dimensional point cloud data is mapped into a two-dimensional depth image for use in convolutional neural network training, including the following steps: Step 3.1, map the 3D image to 2D: Each original point cloud (x, y, z) is mapped to a two-dimensional coordinate (u, v) by the following formula: Where w and h represent the width and height of the projected image; r represents the distance between each point and the radar; f down is the vertical field of view below the LiDAR sensor; f up is the vertical field of view above the LiDAR sensor; f is the overall vertical field of view of the sensor; The 3D point cloud coordinates (x, y, z), intensity values ​​(i) and distance values ​​(r) are stored in the RV image channel respectively, and the final generated [w×h×5] projection image is input into the neural network for learning; Step 3.2: Build the dataset: The annotation system is used to accurately annotate the gantry drainage lines in the two-dimensional image at the pixel level; and a point cloud segmentation dataset is constructed with 70% of the data used for training, 20% of the data used for verification, and 10% of the data used for testing.

6. The method for intelligent safety monitoring of the gantry drainage line of a substation according to claim 1 is characterized in that: In step 4, the lightweight deep learning network is a lightweight semantic segmentation model; the lightweight semantic segmentation model uses MobileNetV3 as the feature extraction network; the lightweight semantic segmentation model integrates global information through the context module, uses the pixel shuffle layer Pixel-Shuffle for upsampling, and combines the feature maps of the encoder and decoder in the decoding stage through the jump connection; after the jump connection, the deep and shallow information are further fused; The lightweight semantic segmentation model introduces a proxy attention AA module, which expands the traditional attention mechanism's quadruple Q, K, V, A by introducing an additional proxy vector A. In the new mechanism, the proxy vector A acts as a proxy for the query vector Q, aggregating information from the key vector K and the value vector V, and feeding the aggregated information back to the query vector Q.

7. The method for intelligent safety monitoring of the gantry drainage line of a substation according to claim 6 is characterized in that: Based on the lightweight semantic segmentation model, the calculation formula of the upsampling method is as follows: PS(T) x,y,c =T [x / r],[y / r],C×mod(x,r)+c Where PS is the periodic pixel shuffle operator; T is the output high-resolution feature map; x, y are the positions of pixels in the output high-resolution feature map; c is the number of channels of the input feature map; T [x / r],[y / r] Indicates that the coordinates x and y of the output feature map are mapped back to the coordinates of the input feature map through integer division operation; C×mod(x,r) performs a modulus operation between x and r to obtain the corresponding channel position; After PS operation, the input is [H×W×r 2 C] will become [rH×rW×C], where r is the upsampling factor; During the training process, an optimized high-resolution feature map can be obtained by learning the weights of r×r Shuffle channels.

8. The method for intelligent safety monitoring of the gantry drainage line of a substation according to claim 1 is characterized by: Based on the lightweight semantic segmentation model, the model training process is as follows: Model training is performed based on the point cloud segmentation dataset constructed in step 3; Use moment estimation optimizer as the training optimization strategy for the overall branch; During the training process, a total of 100 rounds were conducted; the initial learning rate was set to 0.0003, and the learning rate was dynamically adjusted after each round of update, multiplied by 0.9; the model loss was monitored, and the training was stopped early when the loss value did not decrease for 5 consecutive rounds; Through the above training method, multiple experiments are carried out to obtain multiple network weight parameter files; The model segmentation accuracy of different weight parameter files is compared on the test set, and the weight parameters with the highest accuracy are selected as the network parameters of the final model.

9. The method for intelligent safety monitoring of the gantry drainage line of a substation according to claim 1, characterized in that: In step 5, during the data post-processing process, the GPU-based K nearest neighbor (KNN) algorithm is introduced for rapid correction. The KNN algorithm searches for the nearest K points in the Euclidean space and then performs a consensus vote to correct the label of the point. Finally, the segmented pixel points are fitted into drainage lines using a polynomial fitting method.

10. The method for intelligent safety monitoring of the gantry drainage line of a substation according to claim 1, characterized in that: In step 6, the distance between drainage lines is calculated as follows: In the formula, D(A,B) represents the Hausdorf f distance between drainage line set A and drainage line set B; A is the point set of the first drainage line; B is the point set of the second drainage line; a represents any point in set A; b represents any point in set B; d(ab) represents the distance between point a and point b; The threshold is set according to the actual situation. If the Hausdorff distance is less than the threshold, an alarm is triggered to notify relevant personnel to take necessary safety measures.

Images