A substation gantry drainage line safety intelligent monitoring method

Abstract

The application relates to the technical field of power line monitoring, and discloses a gantry frame drainage line safety intelligent monitoring method for a transformer substation, which is based on laser point cloud spherical projection and comprises the following steps: obtaining accurate point cloud information of the gantry frame drainage line, pre-processing the collected point cloud data, projecting the spherical surface of the gantry frame drainage line to a plane to obtain a two-dimensional image, training by using an improved lightweight deep learning network, accurately segmenting pixel points of the drainage line from the point cloud data, performing feature fitting on the drainage line, and calculating the distance between two drainage lines; compared with a method of directly using three-dimensional point cloud data for segmentation, the method reduces the data volume, reduces the labeling difficulty, balances the prediction speed and accuracy, and can realize the function of real-time early warning on an edge device. The method detects the distance between the gantry frame drainage lines, discovers safety hazards in time, triggers an early warning signal, and ensures the safe operation of the transformer substation.

Description

Technical Field

[0001] This invention relates to the field of power line monitoring technology, specifically to a method for intelligent monitoring of the safety of substation gantry feed lines. Background Technology

[0002] The increasing demand for electricity and the expansion of power systems have placed higher demands on their safety and stability. Transformers are crucial components of power systems, undertaking the vital task of voltage transformation; their safe and stable operation directly impacts the overall stability of the power system. While gantry crane conductors have a certain margin, under complex environmental conditions, a short-term misalignment of two-phase conductors can lead to a phase-to-phase short-circuit fault due to short-circuit electrodynamic forces. This can cause the main transformer to trip, disrupting the stability of the power system and jeopardizing 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 guarantee timeliness. Using laser point cloud technology for real-time intelligent safety detection is of great significance for maintaining the safety of substation systems.

[0004] In power line monitoring and early warning systems, an existing technical solution is based on point cloud data processing and deep learning models. This method obtains high-quality point cloud data by performing outlier removal, thinning, and registration on the dataset. Then, based on line detection and region growing segmentation algorithms, the point cloud data is segmented to extract power line point clouds, and finally, the power line point clouds are classified.

[0005] Existing power line fitting techniques are mainly divided into three categories: methods based on raw point clouds, methods based on voxels, and methods based on projection. Methods based on raw point clouds directly use the raw point cloud as input; however, these methods have high computational efficiency and memory requirements, making them difficult to handle large-scale point cloud data. Point cloud voxelization involves segmenting irregular point clouds into regular voxels and then using 3D convolution for feature learning; while this method can clearly encode 3D spatial information, because the input is a fixed-grid voxel, all points within the same voxel have the same semantic label; furthermore, as the grid resolution increases, memory and computational costs grow exponentially, making it impossible to achieve fine-grained inference in large-scale scenes. Projection-based methods project the point cloud onto a plane or curved surface, generating a dense and compact 2D projection map; although this method can flexibly adapt to different scene scales, existing techniques typically increase model complexity to improve segmentation accuracy, making it difficult to achieve a balance between segmentation accuracy and speed. Furthermore, existing techniques are limited to fitting power lines and fail to achieve further early warning functions, leaving a gap in safety early warning for substation gantry line feeders. Summary of the Invention

[0006] To address the aforementioned technical shortcomings and meet the requirements for algorithm accuracy and real-time performance in substation scenarios, this invention proposes a smart safety monitoring method for substation gantry feeder lines. This method is based on spherical projection of laser point clouds, projecting laser point cloud data spherically onto a two-dimensional plane. A real-time semantic segmentation network is used to balance detection speed, detection accuracy, equipment computing power, and storage requirements, thereby improving the segmentation accuracy of gantry feeder line extraction. Furthermore, an early warning module provides real-time warnings for short-circuit conditions in the feeder lines, ensuring the normal operation of the substation.

[0007] The technical solution adopted in this invention is as follows:

[0008] A method for intelligent safety monitoring of substation gantry cable guide lines, based on laser point cloud spherical projection, includes the following steps:

[0009] Step 1. Use lidar equipment to collect 3D point cloud data at the substation site to obtain accurate point cloud information of the gantry cable guide 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, project the spherical projection of the gantry drainage line onto the plane to obtain a two-dimensional image, perform manual annotation, and construct a gantry drainage line dataset;

[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 guide line from the point cloud data, and perform post-processing on the segmentation results;

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

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

[0016] Step 1.1, Installation and parameter settings of the lidar equipment:

[0017] Install lidar equipment in key locations in the substation to ensure full coverage of the gantry cable guide line; set lidar scanning parameters, including scanning frequency, angle range, and resolution, to ensure high-quality and efficient acquisition of point cloud data;

[0018] Step 1.2, Point Cloud Data Acquisition and Environmental Parameter Recording:

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

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

[0021] Gaussian distribution filtering is used for noise reduction. By analyzing the neighborhood of each point, a threshold is determined using the mean and standard deviation of the Gaussian distribution. Points exceeding the threshold are considered noise and are removed.

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

[0023]

[0024] In the formula, Indicates the location The probability density value at that location; This is a parameter for the smoothness of the function; Standardization factor; The base of the natural logarithm; This is the exponential part of the Gaussian function; This represents the expected value of a normal distribution.

[0025] Value Transformation of two-dimensional vectors The two-dimensional form of Gaussian filtering is:

[0026]

[0027] In the formula, The weights representing the probability density values ​​in two-dimensional space; Standardization factor; This is a parameter for the smoothness of the function; The base of the natural logarithm; For the exponential part; It 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 center point in the neighborhood point cloud and substitute it into the Gaussian function to generate a Gaussian template; if the Gaussian template contains small values, normalize the generated Gaussian template so that its upper left corner value is 1; align the center of the Gaussian template with the point cloud matrix to be processed; multiply the corresponding elements and add them together, padding the positions with no elements with zeros;

[0030] The above calculations are performed on each point cloud element to finally generate the denoised output matrix.

[0031] Furthermore, in step 2, the ground filtration process is as follows:

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

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

[0034]

[0035] In the formula, For the plane in Inclination coefficient in the direction; In order to be in Inclination coefficient in the direction; plane in Inclination coefficient in the direction; For constant terms; Let be the coordinate axis variables in three-dimensional space; where, Not both are 0;

[0036] Three points were randomly selected from the substation site cloud. , , The plane parameter values ​​can be solved using the following formula:

[0037]

[0038] In the formula, These are the coefficients of the plane equation; For constant terms;

[0039] Calculate planar distance The calculation formula is as follows:

[0040]

[0041] In the formula, point Distance to the plane; These are the coefficients of the plane equation; The point is Coordinates in three-dimensional space; For constant terms;

[0042] Based on the obtained planar model, the planar distance Points with a distance greater than the threshold are considered outside points, and points with a distance less than the threshold are considered inside points; the threshold is typically between 0.01 and 0.1.

[0043] Update the parameter model based on the number of in-place points, and continue iterating until the specified number of iterations is reached;

[0044] Number of iterations Calculate using the following formula:

[0045]

[0046] In the formula, The ratio of intra-points to the total number of point clouds; This represents the probability that the selected sample point during the iteration process is an inlier. This refers to the number of points required to simulate the plane during the iteration process;

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

[0048] Furthermore, step 3 involves mapping the unstructured 3D point cloud data into a 2D depth image for use in training the convolutional neural network, including the following steps:

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

[0050] Each original point cloud Mapped to two-dimensional coordinates using the following formula :

[0051]

[0052] In the formula, and Indicates the width and height of the projected image; This indicates the distance between each point and the radar; The vertical field of view below the lidar sensor; The vertical field of view above the lidar sensor; This represents the sensor's overall vertical field of view.

[0053] 3D point cloud coordinates Strength value and distance value Stored separately in the RV image channels, and finally generated The projected image is input into the neural network for learning;

[0054] Step 3.2: Construct the dataset:

[0055] A pixel-level precise annotation system was used to annotate the gantry guide lines in the two-dimensional image; and a point cloud segmentation dataset was constructed with 70% of the data used for training, 20% for validation, and 10% 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 the feature extraction network; the lightweight semantic segmentation model integrates global information through the context module, uses the pixel-shuffle layer for upsampling, and combines the feature maps of the encoder and decoder in the decoding stage through skip connections; after the skip connections, the deep and shallow information are further fused.

[0057] The lightweight semantic segmentation model introduces a proxy attention module (AA). By introducing an additional proxy vector A, the quadruple Q, K, V, A of the traditional attention mechanism is extended. 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 for its upsampling method is as follows:

[0059]

[0060] In the formula, It is a periodic pixel shuffle operator; This is the output high-resolution feature map; It represents the position of the pixel in the output high-resolution feature map; The number of channels in the input feature map; This indicates that the coordinates of the output feature map will be determined through integer division. and Map back to the coordinates of the input feature map; exist and Modulo operations are performed between them to obtain the corresponding channel positions;

[0061] go through Operation, input is The size of the feature map will become r is the upsampling factor;

[0062] During training, one can learn... The weights of each Shuffle channel are used to obtain an optimized high-resolution feature map.

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

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

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

[0066] Multiple experiments were conducted using the training method described above to obtain multiple network weight parameter files. The segmentation accuracy of the models with different weight parameter files was compared on the test set, and the weight parameters with the highest accuracy were selected as the network parameters of the final model.

[0067] Furthermore, in step 5, during data post-processing, a GPU-based K-nearest neighbor (KNN) algorithm is introduced for rapid correction. The KNN algorithm corrects the label of a point by searching for the K nearest points in Euclidean space and then performing consensus voting. Finally, the segmented pixels are fitted into a flow line using a polynomial fitting method.

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

[0069]

[0070] In the formula, Represents the set of drainage lines and drainage line collection The Hausdorff distance between them; It is the set of points of the first drainage line; It is the set of points of the second drainage line; Represents a set Any point in it; Represents a set Any point in it; This represents the distance between point a and point b;

[0071] A threshold is set based on the actual situation. If the Hausdorff distance is less than the threshold, an alarm is triggered, and relevant personnel are notified to take necessary safety measures.

[0072] Compared with existing technologies, this intelligent monitoring method for the safety of substation gantry feeder lines has the following advantages:

[0073] This intelligent safety monitoring method for the substation gantry feeder lines is based on laser point cloud spherical projection. By projecting 3D point cloud data spherically onto a 2D plane, compared to directly segmenting 3D point cloud data, it reduces the amount of data, lowers the annotation difficulty, and balances prediction speed and accuracy. This enables real-time early warning functionality when deployed at edge devices. Compared to planar projection, spherical projection better preserves the spatial relationships in the point cloud data, improving segmentation accuracy.

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

[0075] The intelligent safety monitoring method for the substation gantry feeder uses a pixel-washing layer instead of the traditional transposed convolution, achieving a highly efficient and lightweight decoding process, avoiding the checkerboard effect, and improving decoding efficiency and segmentation accuracy.

[0076] In summary, this intelligent safety monitoring method for substation gantry lead-out lines can operate efficiently on edge devices. By detecting the distance between gantry lead-out lines, it can promptly identify safety hazards and trigger early warning signals, ensuring the safe operation of the substation. Attached Figure Description

[0077] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein:

[0078] Figure 1 This is a flowchart of the intelligent safety monitoring method for substation gantry diversion lines according to the present invention;

[0079] Figure 2 This is the judgment logic diagram of the intelligent monitoring method for the safety of substation gantry diversion lines according to the present invention. Detailed Implementation

[0080] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0081] This invention proposes a method for intelligent safety monitoring of substation gantry line guide wires; the process of this method is as follows: Figure 1 As shown, the judgment logic is as follows: Figure 2 As shown, the specific steps include:

[0082] Step 1. Use lidar equipment to collect 3D point cloud data at the substation site to obtain accurate point cloud information of the gantry cable guide; specifically including the following steps:

[0083] Step 1.1, Installation and parameter settings of the lidar equipment:

[0084] Install lidar equipment at key locations in the substation to ensure comprehensive coverage of the gantry cable guide lines. Configure lidar scanning parameters, including scanning frequency, angle range, and resolution, to ensure high-quality and efficient point cloud data acquisition.

[0085] Step 1.2, Point Cloud Data Acquisition and Environmental Parameter Recording:

[0086] 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.

[0087] Step 2. Preprocess the acquired point cloud data, including point cloud denoising, filtering, and ground filtering, to improve data quality and usability; the specific process is as follows:

[0088] Step 2.1, Point Cloud Denoising:

[0089] Due to varying surface roughness of the scanned objects and the influence of external environmental noise, point cloud data may contain small patches of point cloud and discrete points that differ from the main point cloud features and are located far from the main point cloud. This noise is detrimental to the extraction and matching of point cloud features.

[0090] To address this issue, this embodiment employs a Gaussian distribution filtering method to denoise the original point cloud data, thereby improving the accuracy of subsequent registration between laser point clouds and visible light images.

[0091] 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 removes points that exceed the threshold as noise points.

[0092] The one-dimensional form of the Gaussian function is:

[0093]

[0094] In the formula, Indicates the location The probability density value at that location; σ is a parameter for the smoothness of the function; the larger the value of σ, the better the smoothness. The standardization factor is used to ensure that the area of ​​the probability density function is equal to 1. The base of the natural logarithm is approximately 2.71828; The exponential part of the Gaussian function represents the distance from the mean. The ratio of the square of the standard deviation to the square of the mean reflects... How close or far from the mean; The expected value of the normal distribution is represented, which determines the center of symmetry of the distribution function.

[0095] Value Transformation of two-dimensional vectors The two-dimensional form of Gaussian filtering is:

[0096]

[0097] In the formula, The weights representing the probability density values ​​in two-dimensional space; This is a normalization factor used to ensure that the integral of the entire two-dimensional Gaussian function equals 1; This is a parameter for the smoothness of the function; The base of the natural logarithm is approximately 2.71828; The exponential part represents the ratio of the squared distance from the origin (0,0) to the squared standard deviation. It is the square of the Euclidean distance from a point to the origin, reflecting how close the point is to the center point on the plane.

[0098] The two-dimensional Gaussian function described above is used to perform linear convolution on the neighborhood of each point cloud. Points at different locations within the neighborhood are assigned different weights, with points closer to each other having a larger weight and points farther away having a smaller weight.

[0099] The specific calculation process is as follows:

[0100] Calculate the distance from other point clouds in the neighborhood to the neighborhood center, and substitute this distance into the Gaussian function to generate a Gaussian template;

[0101] If the Gaussian template contains small values, it is normalized by normalizing the value at the top left corner of the template to 1.

[0102] Align the center of the Gaussian template with the point cloud matrix to be processed and perform linear convolution operation; that is, multiply the corresponding elements and then add them together, padding the positions with no elements with zeros.

[0103] The above calculations are performed on each point cloud element to finally generate the denoised output matrix.

[0104] Step 2.2, Ground filtration:

[0105] This embodiment uses the RANSAC method to filter out the ground; RANSAC is a statistical probability-based algorithm that, after obtaining the shape of the object in advance, solves the mathematical model parameters representing the object through an iterative method, and finally obtains a point cloud set representing the object.

[0106] The ground is a plane in space, and its equation is expressed as follows:

[0107]

[0108] In the formula, For the plane in Inclination coefficient in the direction; In order to be in Inclination coefficient in the direction; plane in Inclination coefficient in the direction; This is a constant term used to adjust the relative position of the plane with respect to the origin; The coordinate axis variable in three-dimensional space represents the coordinate axis variable in three-dimensional space. , and The coordinate values ​​in the direction; where, They are not both 0.

[0109] Three points were randomly selected from the substation site cloud. , , The plane parameter values ​​can be solved using the following formula:

[0110]

[0111] In the formula, These are the coefficients of the plane equation, representing the components of the plane's normal vector; This is a constant term that controls the height of the plane in space and affects the distance between the plane and the origin.

[0112] Based on the obtained planar model, the planar distance Points with a distance greater than the threshold are considered outside points, and points with a distance less than the threshold are considered inside points; the threshold is typically between 0.01 and 0.1.

[0113]

[0114] In the formula, point Distance to the plane; These are the coefficients of the plane equation; The point is Coordinates in three-dimensional space; For constant terms; For calculation points The algebraic distance to the plane; Calculate the magnitude of the plane normal vector;

[0115] Each iteration calculates the error between the inliers and the parametric model, then updates the data and recalculates the parametric model until the iteration is complete.

[0116] Number of iterations Calculate using the following formula:

[0117]

[0118] In the formula, The ratio of intra-points to the total number of point clouds; This represents the probability that the selected sample point during the iteration process is an inlier. This represents the number of points required to simulate the plane during the iteration process.

[0119] After the iteration is completed, the plane with the most local points is taken as the ground, and the ground points are filtered out from the point cloud data.

[0120] Step 3. After filtering out the ground, project the spherical projection of the gantry drainage line onto the plane to obtain a two-dimensional image, perform manual annotation, and construct a gantry drainage line dataset; the specific process is as follows:

[0121] Step 3.1: Map the 3D image to 2D:

[0122] Unstructured 3D point clouds are spherically projected to obtain the raw depth image of the LiDAR, which can then be fed into a standard convolutional neural network for learning.

[0123] In the 2D projection map RV, each original point cloud Mapped to two-dimensional coordinates using the following formula :

[0124]

[0125] In the formula, and Indicates the width and height of the projected image; The distance between each point and the radar is represented by the following formula: ; The vertical field of view below the lidar sensor; The vertical field of view above the lidar sensor; The overall vertical field of view of the sensor is defined as: .

[0126] To preserve as much of the complete field of view and spatial information as possible from the rotating lidar, the 3D point cloud coordinates... Strength value and distance value Stored separately in the RV image channels, and finally generated The projected image is input into the neural network for learning.

[0127] Step 3.2: Construct the dataset:

[0128] A pixel-level precise annotation system was used to annotate the gantry guide lines in the two-dimensional image; and a point cloud segmentation dataset was constructed with 70% of the data used for training, 20% for validation, and 10% for testing.

[0129] The above steps demonstrate the idea of ​​segmenting the gantry drain line by projecting the spherical laser point cloud onto a plane. This transforms complex 3D data into a 2D plane, which is then input into the segmentation model for training. This reduces the amount of point cloud data, meets the computing power requirements of edge devices and the real-time requirements of substation scenarios, and ensures segmentation accuracy. Projecting the laser point cloud data spherically onto a 2D plane reduces data complexity while meeting real-time requirements. This method not only effectively reduces the amount of data input to the model but also guarantees the accuracy and real-time performance of the segmentation model in substation scenarios.

[0130] Step 4. Based on the constructed gantry drain line dataset, an improved lightweight deep learning network is used for training to enhance the model's accuracy and efficiency; the specific process is as follows:

[0131] Step 4.1, Design a lightweight semantic segmentation model:

[0132] The lightweight semantic segmentation network mentioned in step 4 follows an hourglass architecture of encoder and decoder. 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 shuffling layer for upsampling and fuses the features extracted during downsampling with the same size feature map of the encoder through skip connections. To further utilize spatial detail information, deep and shallow information are fused through a feature fusion module after the skip connections.

[0133] In substation scenarios, foreground objects such as gantry cranes and feed lines exhibit a severe imbalance with background objects, such as substation facilities and buildings. This poses a challenge to traditional semantic segmentation networks during the feature extraction stage. Especially when processing small targets, due to their relatively small pixel count, crucial information is easily lost during downsampling.

[0134] To address this issue, this embodiment introduces an AgentAttention (AA) module. The low-level features of an image contain rich contour, edge, and positional information, which is crucial for accurately identifying the gantry drain line.

[0135] However, traditional convolutional operations may not be able to capture enough contextual information when processing these features, so a proxy attention mechanism is used to enhance the feature extraction process.

[0136] 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.

[0137] This module not only effectively captures the key features of the gantry drain line, but also improves processing speed while maintaining computational efficiency. The surrogate attention mechanism combines the advantages of the Softmax attention mechanism and the linear attention mechanism, possessing low computational complexity and high model expressive power; its characteristics enable the model to extract features more accurately and maintain high computational speed when dealing with small foreground targets in substation scenarios.

[0138] Semantic segmentation networks based on encoder-decoder structures typically use transposed convolutions to upsample feature maps. However, transposed convolutions not only require significant GPU memory during training but also cause a checkerboard effect due to uneven overlap during the mapping process. This embodiment uses a pixel-shuffle layer instead of transposed convolutions for upsampling.

[0139] The Pixel-Shuffle layer generates high-resolution feature maps by randomly arranging the learned feature maps in the channel dimension to the spatial dimension. This process is also known as sub-pixel convolution. Specifically, Pixel-Shuffle treats each pixel of a low-resolution image as multiple sub-pixels and interpolates these sub-pixels across multiple channels according to certain rules, achieving pixel recombination. Sub-pixel convolution periodically activates pixel shuffling operations based on the positions of different sub-pixels.

[0140] The calculation formula is as follows:

[0141]

[0142] In the formula, It is a periodic pixel shuffle operator; This is the output high-resolution feature map; It represents the position of the pixel in the output high-resolution feature map; The number of channels in the input feature map; This indicates that the coordinates of the output feature map will be determined through integer division. and Map back to the coordinates of the input feature map; exist and Modulo operations are performed between them to obtain the corresponding channel positions.

[0143] go through Operation, input is The size of the feature map will become r is the upsampling factor;

[0144] During training, one can learn... The weights of each Shuffle channel are used to obtain an optimized high-resolution feature map.

[0145] Step 4.2, Model Training:

[0146] The laser point cloud gantry drain line detection model needs to be trained before segmentation; the steps are as follows:

[0147] a) Training was performed on the point cloud segmentation dataset constructed in step three, using a moment estimation optimizer as the overall branch training optimization strategy, for a total of 100 training epochs. The learning rate was dynamically adjusted starting from 0.0003, and multiplied by 0.9 after each epoch update. The model loss was calculated after each training epoch, and training was stopped when the loss value did not decrease for 5 consecutive epochs.

[0148] b) Train the model multiple times using the training method in step a), and obtain multiple network weight parameter files for the model through multiple trainings. Compare the segmentation accuracy of the models using different weight parameter files on the test set, and select the network weight parameters with the highest accuracy as the parameters of each network in the model.

[0149] In step 4, a lightweight semantic segmentation model was designed, employing 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 during the inference phase but also maintains high segmentation accuracy, thereby achieving real-time detection capability for gantry drain lines.

[0150] Step 5. Apply the trained deep learning model to accurately segment the pixels of the drainage lines from the point cloud data, and perform post-processing on the segmentation results; the specific process is as follows:

[0151] Step 5.1, Data Post-processing:

[0152] After obtaining the pixel-by-pixel classification results of the projection map, the segmentation results need to be mapped back to the 3D point cloud to obtain point-by-point labels. However, due to the segmentation error of the projection map 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, this invention introduces a GPU-based K-nearest neighbor (KNN) algorithm for fast correction.

[0153] The KNN algorithm corrects the label of a point by searching for the K nearest points in Euclidean space and then performing a consensus vote.

[0154] Step 5.2: Drainage line fitting:

[0155] The segmented pixels are fitted into a flow line using a polynomial fitting method.

[0156] Step 6. Based on the segmentation results of the deep learning model, perform feature fitting on the drainage lines and calculate the distance between the two drainage lines; once the distance is less than the preset threshold, the system will issue an alarm signal in real time.

[0157] Step 6.1, calculate the distance between drainage lines:

[0158] Calculate the distance between the two drain lines, and then calculate the Hausdorff distance between the two wires, which is the minimum of the maximum distances from each drain line to the other. The formula for this calculation is as follows:

[0159]

[0160] In the formula, Represents the set of drainage lines and drainage line collection The Hausdorff distance between them; It is the set of points of the first drainage line; It is the set of points of the second drainage line; Represents a set Any point in it; Represents a set Any point in it; This represents the distance between point a and point b.

[0161] Step 6.2, Judgment and Alarm:

[0162] A threshold is set based on the actual situation. If the Hausdorff distance is less than the threshold, an alarm is triggered, and relevant personnel are notified to take necessary safety measures.

[0163] In summary, this intelligent safety monitoring method for substation gantry feeder lines, combining point cloud data and deep learning models, forms a specialized intelligent safety monitoring method for real-time detection and monitoring of substation gantry feeder lines. It can promptly detect abnormal distances between gantry feeder lines, thereby preventing potential safety risks and ensuring the safe and continuous operation of the substation.

[0164] It will be apparent to those skilled in the art that the present 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 its spirit or essential characteristics. 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. Thus, all variations falling within the meaning and scope of the same elements of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A method for intelligent safety monitoring of substation gantry feed lines, characterized in that: The intelligent safety monitoring method for the substation gantry feeder is based on laser point cloud spherical projection and includes the following steps: Step 1. Use lidar equipment to collect 3D point cloud data at the substation site to obtain accurate point cloud information of the gantry cable guide 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, project the spherical projection of the gantry drainage line onto the plane to obtain a two-dimensional image, perform manual annotation, and construct a gantry drainage line dataset; 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 guide line from the point cloud data, and perform post-processing on the segmentation results; Step 6. Based on the segmentation results of the deep learning model, perform feature fitting on the drainage lines and calculate the distance between the two drainage lines; once the distance is less than the preset threshold, the system will issue an alarm signal in real time. Step 3 includes the following steps: Step 3.1: Map the 3D image to 2D: Each original point cloud Mapped to two-dimensional coordinates using the following formula : ； In the formula, and Indicates the width and height of the projected image; This indicates the distance between each point and the radar; The vertical field of view below the lidar sensor; The vertical field of view above the lidar sensor; The overall vertical field of view of the sensor is defined as: ; 3D point cloud coordinates Strength value and distance value Stored separately in the RV image channels, and finally generated The projected image is input into the neural network for learning; Step 3.2: Construct the dataset: A pixel-level precise annotation system was used to annotate the gantry guide lines in the 2D image; and a point cloud segmentation dataset was constructed with 70% of the data used for training, 20% for validation, and 10% for testing. 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 for upsampling, and combines the feature maps of the encoder and decoder in the decoding stage through skip connections; after the skip connections, the deep and shallow information are further fused. The lightweight semantic segmentation model introduces a proxy attention module (AA). By introducing an additional proxy vector A, the quadruple Q, K, V, A of the traditional attention mechanism is extended. 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.

2. The intelligent monitoring method for the safety of substation gantry feeder lines according to claim 1, characterized in that: Step 1 includes the following steps: Step 1.1, Installation and parameter settings of the lidar equipment: Install lidar equipment in key locations in the substation to ensure full coverage of the gantry cable guide line; set lidar scanning parameters, including scanning frequency, angle range, and resolution, to ensure high-quality and efficient acquisition of point cloud data; Step 1.2, Point Cloud Data Acquisition and Environmental Parameter Recording: For monitoring scenarios, multiple sets of point cloud data are collected, and relevant parameters and environmental information are recorded in detail during the collection process.

3. The intelligent safety monitoring method for substation gantry feeder lines according to claim 1, characterized in that: In step 2, the point cloud denoising process is as follows: Gaussian distribution filtering is used for noise reduction. By analyzing the neighborhood of each point, a threshold is determined using the mean and standard deviation of the Gaussian distribution. Points exceeding the threshold are considered noise and are removed. The one-dimensional form of the Gaussian function is: ； In the formula, Indicates the location The probability density value at that location; This is a parameter for the smoothness of the function; Standardization factor; The base of the natural logarithm; This is the exponential part of the Gaussian function; This represents the expected value of a normal distribution. Value Transformation of two-dimensional vectors The two-dimensional form of Gaussian filtering is: ； In the formula, The weights representing the probability density values ​​in two-dimensional space; Standardization factor; This is a parameter for the smoothness of the function; The base of the natural logarithm; For the exponential part; It is the square of the Euclidean distance from a point to the origin; The calculation process is as follows: Calculate the distance from the center point in the neighborhood point cloud and substitute it into the Gaussian function to generate a Gaussian template; if the Gaussian template contains small values, normalize the generated Gaussian template so that its upper left corner value is 1; align the center of the Gaussian template with the point cloud matrix to be processed; multiply the corresponding elements and add them together, padding the positions with no elements with zeros; The above calculations are performed on each point cloud element to finally generate the denoised output matrix.

4. The intelligent monitoring method for safety of substation gantry feeder lines according to claim 1, characterized in that: In step 2, the ground filtration process is as follows: The RANSAC method is used to filter out the ground. After obtaining the shape of the object in advance, the mathematical model parameters representing the object are solved iteratively to finally obtain the point cloud set representing the object. The ground is a plane in space, and its equation is expressed as follows: ； In the formula, For the plane in Inclination coefficient in the direction; In order to be in Inclination coefficient in the direction; plane in Inclination coefficient in the direction; For constant terms; Let be the coordinate axis variables in three-dimensional space; where, Not both are 0; Three points were randomly selected from the substation site cloud. , , The plane parameter values ​​can be solved using the following formula: ； In the formula, These are the coefficients of the plane equation; For constant terms; Calculate planar distance The calculation formula is as follows: ； In the formula, point Distance to the plane; These are the coefficients of the plane equation; The point is Coordinates in three-dimensional space; For constant terms; Based on the obtained planar model, the planar distance Points with a distance greater than the threshold are considered outside points, and points with a distance less than the threshold are considered inside points; the threshold is typically between 0.01 and 0.

1. Update the parameter model based on the number of in-place points, and continue iterating until the specified number of iterations is reached; Number of iterations Calculate using the following formula: ； In the formula, The ratio of intra-points to the total number of point clouds; This represents the probability that the selected sample point during the iteration process is an inlier. This refers to the number of points required to simulate the plane during the iteration process; After the iteration is completed, the plane with the most local points is taken as the ground, and the ground points are filtered out from the point cloud data.

5. The intelligent monitoring method for the safety of substation gantry feeder lines according to claim 1, characterized in that: The calculation formula for the upsampling method based on the lightweight semantic segmentation model is as follows: ； In the formula, It is a periodic pixel shuffle operator; This is the output high-resolution feature map; It represents the position of the pixel in the output high-resolution feature map; The number of channels in the input feature map; This indicates that the coordinates of the output feature map will be determined through integer division. and Map back to the coordinates of the input feature map; exist and Modulo operations are performed between them to obtain the corresponding channel positions; go through Operation, input is The size of the feature map will become r is the upsampling factor; During the training process, through learning The weights of each Shuffle channel are used to obtain an optimized high-resolution feature map.

6. The intelligent safety monitoring method for substation gantry feeder lines according to claim 1, characterized in that: The training process for this lightweight semantic segmentation model is as follows: Model training is performed based on the point cloud segmentation dataset constructed in step three; The training optimization strategy uses the moment estimation optimizer as the overall branch; The training process consists of 100 rounds; the initial learning rate is set to 0.0003, and the learning rate is dynamically adjusted by multiplying by 0.9 after each round of updates; the model loss is monitored, and training is stopped early if the loss value does not decrease for 5 consecutive rounds. Multiple experiments were conducted using the above training method to obtain multiple network weight parameter files; The model segmentation accuracy of different weight parameter files was compared on the test set, and the weight parameters with the highest accuracy were selected as the network parameters of the final model.

7. The intelligent monitoring method for safety of substation gantry feeder lines according to claim 1, characterized in that: In step 5, during data post-processing, a GPU-based K-nearest neighbor (KNN) algorithm is introduced for rapid correction. The KNN algorithm corrects the label of a point by searching for the K nearest points in Euclidean space and then performing consensus voting. Finally, the segmented pixels are fitted into a flow line using a polynomial fitting method.

8. The intelligent monitoring method for safety of substation gantry feeder lines according to claim 1, characterized in that: In step 6, the formula for calculating the distance between drainage lines is as follows: ； In the formula, Represents the set of drainage lines and drainage line collection The Hausdorff distance between them; It is the set of points of the first drainage line; It is the set of points of the second drainage line; Represents a set Any point in it; Represents a set Any point in it; This represents the distance between point a and point b; A threshold is set based on the actual situation. If the Hausdorff distance is less than the threshold, an alarm is triggered, and relevant personnel are notified to take necessary safety measures.

Images