Transformer substation portal frame drainage wire safety intelligent monitoring method based on point cloud semantic segmentation technology

Abstract

The invention relates to the technical field of power equipment monitoring, provides a transformer substation portal frame drainage wire safety intelligent monitoring method based on a point cloud semantic segmentation technology, designs a point cloud semantic segmentation model with higher precision, and introduces an attention mechanism and a feature extraction module into a traditional coding-decoding structure. The extraction of different network layer features is enhanced, and the requirement of a substation scene for segmentation precision is better met. By calculating the Hausdorff distance, the system can monitor the position and the state of the drainage wire in time and give an alarm. Meanwhile, Gaussian filtering and RANSAC methods are adopted, the adaptability of the system to environmental noise and complex terrains is improved, the model is optimized through multi-feature extraction and an attention mechanism, efficient segmentation of complex scenes is achieved, and the generalization ability of the model is enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of power equipment monitoring, and in particular to a method for intelligently monitoring the safety of a substation gantry drainage line based on point cloud semantic segmentation technology. 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. In power systems, gantry drainage wires are structural components that connect and support high-voltage conductors and are used to introduce current from the main transmission line to substation equipment. The gantry provides physical support to prevent the conductors from short-circuiting or contacting other objects, thereby ensuring the safety and stability of the power system. There is a certain margin between the gantry drainage wires, but under complex environmental conditions, the two-phase drainage wires will deviate in a short period of time, which will cause phase-to-phase short-circuit faults due to the action of short-circuit electric power, causing the main transformer to trip, destroying the stability of the power system, and endangering the safe and stable operation of the power grid. However, traditional monitoring and maintenance methods often rely on manual inspections, which are inefficient, costly, and prone to missed detection and misjudgment.

[0003] To this end, the prior art introduces an intelligent monitoring method based on point cloud data processing and deep learning models. This method obtains high-quality point cloud data by removing outliers, thinning and registering, and uses line detection and region growing algorithms for segmentation, thereby extracting and classifying power line point clouds. In the prior art, power line fitting mainly includes three methods: based on original point cloud, voxelization and projection method. Although the voxelization method can encode three-dimensional information, it requires a fixed grid, resulting in a surge in memory and computation, making it difficult to achieve fine-grained reasoning in large-scale scenes. The projection method adapts to different scenes by generating two-dimensional projection maps, but increases the complexity of the model to improve accuracy, making it difficult to balance segmentation accuracy and speed, and is easily affected by occlusion. The method based on the original point cloud requires high computing efficiency and memory, and is difficult to process large-scale data. In addition, these methods are prone to losing object boundary details during the segmentation process, affecting the accuracy of power line segmentation. Summary of the invention

[0004] The purpose of the present invention is to solve the shortcomings of the prior art and provide a substation gantry drainage line safety intelligent monitoring method based on point cloud semantic segmentation technology to accurately identify and segment the drainage line, ensure the high accuracy of the monitoring results, and increase the reliability of detection.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions:

[0006] A method for intelligent safety monitoring of a substation gantry drainage line based on point cloud semantic segmentation technology comprises the following steps:

[0007] 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;

[0008] 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;

[0009] Step 3: Use the annotation platform to manually annotate the point cloud data and construct a gantry drainage line dataset;

[0010] Step 4: Based on the gantry drainage line dataset constructed in step 3, train the point cloud segmentation model;

[0011] Step 5: predict the point cloud data based on the point cloud segmentation model trained in step 4, and fit the drainage line points into drainage lines;

[0012] Step 6: Calculate the distance between the two drainage lines. Once the distance is less than the preset threshold, an alarm signal will be issued.

[0013] Preferably, in step 1, the laser radar equipment is installed, and the laser radar equipment is installed in the substation to fully cover the position of the gantry drainage line; the scanning frequency, angle range and resolution of the laser radar are set; and multiple groups of point cloud data are collected for the monitoring scene.

[0014] Preferably, in step 2, the point cloud data is denoised using a Gaussian distribution filtering method. By performing statistical analysis on the neighborhood of each point, the threshold is determined using the mean and standard deviation of the Gaussian distribution, and points exceeding the threshold are regarded as noise points and removed. The one-dimensional form of the Gaussian function f(x) is:

[0015]

[0016] Convert the value x to a two-dimensional vector (x, y). The two-dimensional form of the Gaussian filter f(x, y) is:

[0017]

[0018] In the formula, μ is the mean of the Gaussian distribution, which determines the symmetry center of the distribution function; σ is the function smoothness parameter, which controls the width of the distribution;

[0019] 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.

[0020] Preferably, in step 2, the RANSAC method is used to filter the ground, and its equation is expressed as follows:

[0021] Axx+Bxy+Cxz+D=0

[0022] Where A, B, C are the parameters of the plane equation, which are not 0 at the same time and define the direction of the plane; D is the constant term of the plane equation;

[0023] 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:

[0024]

[0025] 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.

[0026]

[0027] Each iteration calculates the error between the local point and the parameter model, then updates the data and calculates the parameter model again until the iteration is completed. The number of iterations k is calculated by the following formula:

[0028]

[0029] Among them, 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 in the iteration process are in-place points, and n is the number of points required to simulate the plane in the iteration process;

[0030] 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.

[0031] Preferably, in step 3, the filtered gantry flow line point cloud data is manually annotated accurately at the pixel level using an annotation system, and a point cloud segmentation data set is constructed in a ratio of 70% for training, 20% for verification, and 10% for testing.

[0032] Preferably, the training of the point cloud segmentation model in step 4 comprises the following steps:

[0033] Step 4.1, design a point cloud segmentation model, using an encoder-decoder structure. The encoder uses the MS-KPConv module to extract multi-scale features, and then uses the attention module to focus the extracted features on effective information. In the decoder, linear interpolation is used for upsampling, and 1*1 convolution is used to predict semantic labels. Finally, the fully connected layer is used to classify each point in the point cloud.

[0034] The MS-KPConv module defines neighbor points as spherical regions, and convolves the features of the center point and its neighbor points through the "kernel point convolution" operation. For each point P i , construct a star region with a radius of r, including all the stars at distance P i Points that do not exceed the radius constitute the neighbor point set N of the point i ; Calculate each point P ij To the center point P i The distance x ij ; Use the kernel function h(x ij ,ξ m ), the feature vector f of each neighbor point ij Multiply by the kernel function value, and then perform weighted summation to obtain a new eigenvector H m , the kernel function formula is as follows:

[0035] h(x ij ,ξ m )=max(0,1-|x ij -ξ m |)

[0036] By using different sizes of core ξ m Features at different scales can be extracted, and feature vectors H at different scales can be m Perform weighted summation to obtain the final eigenvector f xi ; The formula for feature aggregation is as follows:

[0037]

[0038] W m To convert feature vectors H of different scales m The weight matrix for weighted summation;

[0039] Attention module, introduces channel affinity attention module CAA for high-level channel features;

[0040] The CAA module includes: a compact channel comparator CCC, which is used to simplify the context in each channel and reduce the computational cost; a channel affinity estimator CAE, which is used to adjust the weights through the similarity matrix S to avoid redundancy; and a residual connection, which is used to improve the feature representation output formula by learning the weight α:

[0041] For a given d-dimensional feature map F N×d The CCC module passes through each channel vector c i The shared MLP operated on the above to simplify the context in each channel; replace the N original points with a smaller number of N′=N / ratio, with a ratio ratio>1; use two MLPM q (·) and M k(·) respectively process c i , that is, q i =M q (c i ), and k i =M k (c i ), Q i and k i They are respectively composed of query matrix Q = [q1, q2, ..., q d ] and the key matrix K = [k1, k2, …, k d ], and then calculate the similarity matrix S d×d =Q T K, whose element value in row i and column j represents F N×d The similarity between the i-th channel and the j-th channel;

[0042] The main process of the CAE module is as follows: first, select the maximum similarity in each column of the similarity matrix S, and then expand it to the same size as S; after subtracting the original S, channels with higher similarity will have lower weights, thereby avoiding aggregation of similar and redundant information:

[0043]

[0044] Among them A d×d is the weight matrix; To expand the dimension operation; max d→1 S is the maximum similarity of each column; softmax normalization function;

[0045] Using another MLPM v (·)get Let V = [v1, v2, …, v d ] is the value matrix;

[0046] Using the residual connection method and learning a weight α to simplify learning, the final output is:

[0047] F′ N×d =CAA(F)=F+α·V·A

[0048] where α is the weight of learning; F′ N×d is the final output feature map; F is the original feature map;

[0049] The CAA module can reduce the amount of computation while retaining important feature information through channel-level feature comparison, thereby improving the segmentation performance of multi-scale objects. This mechanism effectively captures the semantic differences in feature maps from different network layers by iteratively aggregating features from adjacent layers, thereby improving feature representation and segmentation performance.

[0050] Step 4.2, model training:

[0051] Use the point cloud segmentation dataset constructed in step 3 for training;

[0052] a) The moment estimation optimizer was used for 100 rounds of training. The initial learning rate was dynamically adjusted from 0.0003. After each round of update, the learning rate was multiplied by 0.9. The model loss was calculated after each round of training. When the loss value did not decrease for 5 consecutive rounds, the training was stopped.

[0053] b) training the model multiple times using the training method in step a), obtaining multiple network weight parameter files of the model through multiple training, comparing the segmentation accuracy of the models using different weight parameter files under the test set, and selecting the network weight parameters with the highest accuracy as the parameters of each network in the model;

[0054] Model evaluation indicators: Overall accuracy (OA), average recall (mRecall) and average intersection over union (mIoU) are selected as evaluation indicators; overall accuracy represents the proportion of correctly predicted categories to the overall input categories, which is used to evaluate the accuracy of overall classification; average recall represents the average recall of each category; average intersection over union represents the intersection of the true value and the predicted value, which is used to evaluate the accuracy of segmentation of all object categories. The specific formula is as follows:

[0055]

[0056]

[0057] In the formula, TP, FP, TN, and FN represent true positive, false positive, true negative, and false negative, respectively. class Represents the number of categories in the dataset.

[0058] Preferably, in step 5, data post-processing is performed, and the classified gantry drainage line point cloud contains multiple drainage lines, and a single drainage line needs to be segmented; multiple drainage lines are divided into a single drainage line using the density-based noisy spatial algorithm DBSCAN; and drainage line fitting is then performed, and the catenary equation is introduced to fit the gantry drainage line. The specific formula is as follows:

[0059]

[0060] Where a is the shape parameter of the control curve; μ, b, and c are the offset and displacement parameters.

[0061] Preferably, in step 6, the distance between the two drainage wires is calculated, and the Hausdorff distance between the two wires is calculated, that is, the minimum value of the maximum distance from each drainage wire to another drainage wire, and the calculation formula is as follows:

[0062]

[0063] Where D(A,B) is the Hausdorff distance between two sets A and B; a and b are points in sets A and B; d(a,b) is the distance between points a and b;

[0064] 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.

[0065] The present invention discloses a method for intelligently monitoring the safety of a substation gantry drainage line based on point cloud semantic segmentation technology, which has the following beneficial effects.

[0066] The present invention designs a more accurate point cloud semantic segmentation model, introduces an attention mechanism and a feature extraction module into the traditional encoding-decoding structure, strengthens the extraction of features from different network layers, and better meets the needs of the substation scene for segmentation accuracy. By calculating the Hausdorff distance, the system can monitor the position and status of the drainage line in a timely manner and issue an alarm. At the same time, the Gaussian filter and RANSAC method are used to improve the system's adaptability to environmental noise and complex terrain. Through multiple feature extraction and attention mechanisms, the model is optimized to achieve efficient segmentation of complex scenes and enhance the generalization ability of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 It is a flow chart of a method for intelligent safety monitoring of a substation gantry drainage line based on point cloud semantic segmentation technology of the present invention;

[0068] Figure 2 This is a schematic diagram of the process in Example 2 of the present invention; DETAILED DESCRIPTION

[0069] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0070] In the description of the present invention, it is necessary to understand that the terms "upper", "lower", "front", "back", "left", "right", "top", "bottom", "inside", "outside", etc., indicating the orientation or position relationship are based on the orientation or position relationship shown in the drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the present invention.

[0071] Example 1

[0072] like Figure 1 As shown, a method for intelligent safety monitoring of substation gantry drainage lines based on point cloud semantic segmentation technology includes the following steps:

[0073] 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;

[0074] Preferably, in step 1, the laser radar equipment is installed, and the laser radar equipment is installed in the substation to fully cover the position of the gantry drainage line; the scanning frequency, angle range and resolution of the laser radar are set; and multiple groups of point cloud data are collected for the monitoring scene.

[0075] 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;

[0076] Preferably, in step 2, the Gaussian distribution filtering method is used to denoise the point cloud data. By performing statistical analysis on the neighborhood of each point, the threshold is determined using the mean and standard deviation of the Gaussian distribution, and points exceeding the threshold are regarded as noise points and removed. The one-dimensional form of the Gaussian function f(x) is:

[0077]

[0078] Convert the value x to a two-dimensional vector (x, y). The two-dimensional form of the Gaussian filter f(x, y) is:

[0079]

[0080] In the formula, μ is the mean of the Gaussian distribution, which determines the symmetry center of the distribution function; σ is the function smoothness parameter, which controls the width of the distribution;

[0081] 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. The specific calculation process is as follows:

[0082] 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; if the Gaussian template contains a small value, normalize it and normalize the value in the upper left corner of the template to 1; 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 without elements; perform the above calculations for each point cloud element, and finally generate a denoised output matrix.

[0083] It is worth noting that due to the different roughness of the surface of the scanned object and the influence of noise in the external environment, the point cloud data may have 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. Therefore, the present invention uses a Gaussian distribution filtering method to denoise the original point cloud data to improve the accuracy of subsequent laser point cloud and visible light image registration.

[0084] Preferably, the RANSAC method is used in step 2 to filter the ground. RANSAC is an algorithm based on statistical probability. 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 equation is expressed as follows:

[0085] Axx+Bxy+Cxz+D=0

[0086] Where A, B, C are the parameters of the plane equation, which are not 0 at the same time and define the direction of the plane; D is the constant term of the plane equation;

[0087] 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:

[0088]

[0089] 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.

[0090]

[0091] Each iteration calculates the error between the local point and the parameter model, then updates the data and calculates the parameter model again until the iteration is completed. The number of iterations k is calculated by the following formula:

[0092]

[0093] Among them, 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 in the iteration process are in-place points, and n is the number of points required to simulate the plane in the iteration process;

[0094] 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.

[0095] Step 3: Use the annotation platform to manually annotate the point cloud data and construct a gantry drainage line dataset;

[0096] Preferably, in step 3, the filtered gantry flow line point cloud data is manually annotated accurately at the pixel level using an annotation system, and a point cloud segmentation data set is constructed in a ratio of 70% for training, 20% for verification, and 10% for testing.

[0097] Step 4: Based on the gantry drainage line dataset constructed in step 3, train the point cloud segmentation model;

[0098] Preferably, the training of the point cloud segmentation model in step 4 includes the following steps:

[0099] Step 4.1, design a point cloud segmentation model, using an encoder-decoder structure. The encoder uses the MS-KPConv module to extract multi-scale features, and then uses the attention module to focus the extracted features on effective information. In the decoder, linear interpolation is used for upsampling, and 1*1 convolution is used to predict semantic labels. Finally, the fully connected layer is used to classify each point in the point cloud.

[0100] The MS-KPConv module defines neighbor points as spherical regions, and convolves the features of the center point and its neighbor points through the "kernel point convolution" operation. For each point P i , construct a star region with a radius of r, including all the stars at distance P i Points that do not exceed the radius constitute the neighbor point set N of the point i ; Calculate each point P ij To the center point P i The distance x ij ; Use the kernel function h(x ij ,ξ m ), the feature vector f of each neighbor point ij Multiply by the kernel function value, and then perform weighted summation to obtain a new eigenvector H m , the kernel function formula is as follows:

[0101] h(x ij ,ξ m )=max(0,1-|x ij -ξ m |)

[0102] By using different sizes of core ξ m Features at different scales can be extracted, and feature vectors H at different scales can be m Perform weighted summation to obtain the final eigenvector f xi ; The formula for feature aggregation is as follows:

[0103]

[0104] W m To convert feature vectors H of different scales m The weight matrix for weighted summation;

[0105] Attention module, introduces channel affinity attention module CAA for high-level channel features;

[0106] The CAA module includes: a compact channel comparator CCC, which is used to simplify the context in each channel and reduce the computational cost; a channel affinity estimator CAE, which is used to adjust the weights through the similarity matrix S to avoid redundancy; and a residual connection, which is used to improve the feature representation output formula by learning the weight α:

[0107] For a given d-dimensional feature map F N×d The CCC module passes through each channel vector c i The shared MLP operated on the above to simplify the context in each channel; replace the N original points with a smaller number of N′=N / ratio, with a ratio ratio>1; use two MLPM q (·) and M k (·) respectively process c i , that is, q i =M q (c i ), and k i =M k (c i ), Q i and k i They are respectively composed of query matrix Q = [q1, q2, ..., q d ] and the key matrix K = [k1, k2, …, k d ], and then calculate the similarity matrix S d×d =Q T K, whose element value in row i and column j represents F N×d The similarity between the i-th channel and the j-th channel;

[0108] The main process of the CAE module is as follows: first, select the maximum similarity in each column of the similarity matrix S, and then expand it to the same size as S; after subtracting the original S, channels with higher similarity will have lower weights, thereby avoiding aggregation of similar and redundant information:

[0109]

[0110] Among them A d×d is the weight matrix; To expand the dimension operation; max d→1 S is the maximum similarity of each column; softmax normalization function;

[0111] Using another MLPM v (·)get Let V = [v1, v2, …, v d ] is the value matrix;

[0112] Using the residual connection method and learning a weight α to simplify learning, the final output is:

[0113] F′ N×d =CAA(F)=F+α·V·A

[0114] where α is the weight of learning; F′ N×d is the final output feature map; F is the original feature map;

[0115] The CAA module can reduce the amount of computation while retaining important feature information through channel-level feature comparison, thereby improving the segmentation performance of multi-scale objects. This mechanism effectively captures the semantic differences in feature maps from different network layers by iteratively aggregating features from adjacent layers, thereby improving feature representation and segmentation performance.

[0116] Step 4.2, model training:

[0117] Use the point cloud segmentation dataset constructed in step 3 for training;

[0118] a) The moment estimation optimizer was used for 100 rounds of training. The initial learning rate was dynamically adjusted from 0.0003. After each round of update, the learning rate was multiplied by 0.9. The model loss was calculated after each round of training. When the loss value did not decrease for 5 consecutive rounds, the training was stopped.

[0119] b) training the model multiple times using the training method in step a), obtaining multiple network weight parameter files of the model through multiple training, comparing the segmentation accuracy of the models using different weight parameter files under the test set, and selecting the network weight parameters with the highest accuracy as the parameters of each network in the model;

[0120] Model evaluation indicators: Overall accuracy (OA), average recall (mRecall) and average intersection over union (mIoU) are selected as evaluation indicators; overall accuracy represents the proportion of correctly predicted categories to the overall input categories, which is used to evaluate the accuracy of overall classification; average recall represents the average recall of each category; average intersection over union represents the intersection of the true value and the predicted value, which is used to evaluate the accuracy of segmentation of all object categories. The specific formula is as follows:

[0121]

[0122] In the formula, TP, FP, TN, and FN represent true positive, false positive, true negative, and false negative, respectively. class Represents the number of categories in the dataset.

[0123] Step 5: predict the point cloud data based on the point cloud segmentation model trained in step 4, and fit the drainage line points into drainage lines;

[0124] As a preference, in step 5, data post-processing is now performed, and the classified gantry drainage line point cloud contains multiple drainage lines, and a single drainage line needs to be segmented; and on the vertical plane where the drainage line is located, the point cloud has obvious aggregation, and there is a certain spatial distance and interval between different conductors, so the density-based noisy spatial algorithm DBSCAN can be used to divide multiple drainage lines into a single drainage line; and then the drainage line is fitted. On the XOZ projection plane, the gantry drainage line is in the form of a catenary, so the present invention introduces a catenary equation to fit the gantry drainage line, and the specific formula is as follows:

[0125]

[0126] Where a is the shape parameter of the control curve; μ, b, and c are the offset and displacement parameters.

[0127] Step 6: Calculate the distance between the two drainage lines. Once the distance is less than the preset threshold, an alarm signal will be issued.

[0128] Preferably, in step 6, the distance between the two drainage wires is calculated, and the Hausdorff distance between the two wires is calculated, that is, the minimum value of the maximum distance from each drainage wire to another drainage wire, and the calculation formula is as follows:

[0129]

[0130] Where D(A,B) is the Hausdorff distance between two sets A and B; a and b are points in sets A and B; d(a,b) is the distance between points a and b;

[0131] 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.

[0132] Example 2

[0133] Based on Example 1, Figure 2 As shown, this embodiment provides a method for intelligently monitoring the safety of a substation gantry drainage line using the point cloud segmentation model obtained in Embodiment 1, including the following steps:

[0134] 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;

[0135] 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;

[0136] Step 3: Use the point cloud segmentation model to segment the drainage line pixels;

[0137] Step 4, determine whether the gantry drainage line is detected; if the gantry drainage line is not detected, repeat steps 1-3; if the gantry drainage line is detected, predict the point cloud data based on the point cloud segmentation model, and fit the drainage line points into the drainage line;

[0138] Step 5: Calculate the distance between the two drainage lines. If the distance is less than the preset threshold, an alarm signal will be issued; if the distance is greater than the preset threshold, repeat steps 1-5.

[0139] The above are only preferred specific embodiments of the present invention, but the protection scope of the present invention is not limited thereto. The substitution may be a substitution of a part of the structure, device, method step, or a complete technical solution. The equivalent replacement or change according to the technical solution and the inventive concept of the present invention shall be covered within the protection scope of the present invention.

Claims

1. A method for intelligent safety monitoring of substation gantry drainage lines based on point cloud semantic segmentation technology, characterized in that: The steps include: 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: Use the annotation platform to manually annotate the point cloud data and construct a gantry drainage line dataset; Step 4: Based on the gantry drainage line dataset constructed in step 3, train the point cloud segmentation model; Step 5: Predict the point cloud data based on the point cloud segmentation model trained in step 4, and fit the drainage line points into drainage lines; Step 6: Calculate the distance between the two drainage lines. Once the distance is less than the preset threshold, an alarm signal will be issued.

2. The method for intelligent safety monitoring of substation gantry drainage lines based on point cloud semantic segmentation technology according to claim 1 is characterized in that: In the step 1, the laser radar equipment is installed, and the laser radar equipment is installed in the substation to fully cover the position of the gantry drainage line; the scanning frequency, angle range and resolution of the laser radar are set; and multiple groups of point cloud data are collected for the monitoring scene.

3. The method for intelligent safety monitoring of substation gantry drainage lines based on point cloud semantic segmentation technology according to claim 1 is characterized in that: In step 2, the Gaussian distribution filtering method is used to denoise the point cloud data. By performing statistical analysis on the neighborhood of each point, the threshold is determined using the mean and standard deviation of the Gaussian distribution, and points exceeding the threshold are regarded as noise points and removed. The one-dimensional form of the Gaussian function f(x) is: Convert the value x to a two-dimensional vector (x, y). The two-dimensional form of the Gaussian filter f(x, y) is: In the formula, μ is the mean of the Gaussian distribution, which determines the symmetry center of the distribution function; σ is the function smoothness parameter, which controls the width of the distribution; 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.

4. The method for intelligent safety monitoring of substation gantry drainage lines based on point cloud semantic segmentation technology according to claim 1 or 3, characterized in that: In step 2, the RANSAC method is used to filter the ground, and its equation is expressed as follows: Axx+Bxy+Cxz+D=0 Where A, B, C are the parameters of the plane equation, which are not 0 at the same time and define the direction of the plane; D is the constant term of the plane equation; 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: 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. Each iteration calculates the error between the local point and the parameter model, then updates the data and calculates the parameter model again until the iteration is completed. The number of iterations k is calculated by the following formula: Among them, 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 in the iteration process are in-place points, and n is the number of points required to simulate the plane in the iteration process; 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 substation gantry drainage lines based on point cloud semantic segmentation technology according to claim 1, characterized in that: In step 3, the filtered gantry flow line point cloud data is manually annotated at the pixel level using an annotation system, and a point cloud segmentation data set is constructed in a ratio of 70% for training, 20% for verification, and 10% for testing.

6. The method for intelligent safety monitoring of substation gantry drainage lines based on point cloud semantic segmentation technology according to claim 1, characterized in that: The training of the point cloud segmentation model in step 4 includes the following steps: Step 4.1, design a point cloud segmentation model, using an encoder-decoder structure. The encoder uses the MS-KPConv module to extract multi-scale features, and then uses the attention module to focus the extracted features on effective information. In the decoder, linear interpolation is used for upsampling, and 1*1 convolution is used to predict semantic labels. Finally, the fully connected layer is used to classify each point in the point cloud. The MS-KPConv module defines neighbor points as spherical regions, and convolves the features of the center point and its neighbor points through the "kernel point convolution" operation. For each point P i , construct a star region with a radius of r, including all the stars at distance P i Points that do not exceed the radius constitute the neighbor point set N of the point i ; Calculate each point P ij To the center point P i The distance x ij ; Use the kernel function h(x ik ,ξ m ), the feature vector f of each neighbor point ij Multiply by the kernel function value, and then perform weighted summation to obtain a new eigenvector H m , the kernel function formula is as follows: h(x ij ,x m )=max(0.1-|x ij -x m |) By using different sizes of core ξ m Features at different scales can be extracted, and feature vectors H at different scales can be m Perform weighted summation to obtain the final eigenvector f xi ; The formula for feature aggregation is as follows: W m To convert feature vectors H of different scales m The weight matrix for weighted summation; Attention module, introduces channel affinity attention module CAA for high-level channel features; The CAA module includes: a compact channel comparator CCC, which is used to simplify the context in each channel and reduce the computational cost; a channel affinity estimator CAE, which is used to adjust the weights through the similarity matrix S to avoid redundancy; and a residual connection, which is used to improve the feature representation output formula by learning the weight α: For a given d-dimensional feature map F N×d The CCC module is implemented by adding a vector c in each channel. i The shared MLP operated on the above to simplify the context in each channel; replace the N original points with a smaller number of N′=N / ratio, with a ratio ratio>1; use two MLPM q (·) and M k (·) respectively process c i , that is, q i =M q (c i ), and k i =M k (c i ), Q i and k i They are respectively composed of query matrix Q = [q1, q2, ..., q d ] and the key matrix K = [k1, k2, …, k d ], and then calculate the similarity matrix S d×d =Q T K, whose element value in row i and column j represents F N×d The similarity between the i-th channel and the j-th channel; The main process of the CAE module is as follows: first, select the maximum similarity in each column of the similarity matrix S, and then expand it to the same size as S; after subtracting the original S, channels with higher similarity will have lower weights, thereby avoiding aggregation of similar and redundant information: Among them A d×d is the weight matrix; To expand the dimension operation; max d→1 S is the maximum similarity of each column; softmax normalization function; Using another MLPM v (·)get Let V = [v1, v2, …, v d ] is the value matrix; Using the residual connection method and learning a weight α to simplify learning, the final output is: F′ N×d =CAA(F)=F+α·V·A Where α is the weight of learning; F′ N×d is the final output feature map; F is the original feature map; The CAA module can reduce the amount of computation while retaining important feature information through channel-level feature comparison, thereby improving the segmentation performance of multi-scale objects. This mechanism effectively captures the semantic differences in feature maps from different network layers by iteratively aggregating features from adjacent layers, thereby improving feature representation and segmentation performance. Step 4.2, model training: Use the point cloud segmentation dataset constructed in step 3 for training; a) The moment estimation optimizer was used for 100 rounds of training. The initial learning rate was dynamically adjusted from 0.0003. After each round of update, the learning rate was multiplied by 0.

9. The model loss was calculated after each round of training. When the loss value did not decrease for 5 consecutive rounds, the training was stopped. b) training the model multiple times using the training method in step a), obtaining multiple network weight parameter files of the model through multiple training, comparing the segmentation accuracy of the models using different weight parameter files under the test set, and selecting the network weight parameters with the highest accuracy as the parameters of each network in the model; Model evaluation indicators: Overall accuracy (OA), average recall (mRecall) and average intersection over union (mIoU) are selected as evaluation indicators; overall accuracy represents the proportion of correctly predicted categories to the overall input categories, which is used to evaluate the accuracy of overall classification; average recall represents the average recall of each category; average intersection over union represents the intersection of the true value and the predicted value, which is used to evaluate the accuracy of segmentation of all object categories. The specific formula is as follows: In the formula, TP, FP, TN, and FN represent true positive, false positive, true negative, and false negative, respectively. class Represents the number of categories in the dataset.

7. The method for intelligent safety monitoring of substation gantry drainage lines based on point cloud semantic segmentation technology according to claim 1, characterized in that: In step 5, data post-processing is now performed. The classified gantry drainage line point cloud contains multiple drainage lines, and a single drainage line needs to be segmented. The multiple drainage lines are divided into a single drainage line using the density-based noisy spatial algorithm DBSCAN. Then, drainage line fitting is performed, and the catenary equation is introduced to fit the gantry drainage line. The specific formula is as follows: Where a is the shape parameter of the control curve; μ, b, and c are the offset and displacement parameters.

8. The method for intelligent safety monitoring of substation gantry drainage lines based on point cloud semantic segmentation technology according to claim 1, characterized in that: In step 6, the distance between the two drainage wires is calculated, and the Hausdorff distance between the two wires is calculated, that is, the minimum value of the maximum distance from each drainage wire to another drainage wire, and the calculation formula is as follows: Where D(A,B) is the Hausdorff distance between two sets A and B; a and b are points in sets A and B; d(a,b) is the distance between points a and 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