A method for establishing a high-voltage line database based on machine learning

Abstract

The present invention discloses a method for establishing a high-voltage line database based on machine learning, including step 1: collecting high-voltage line towers in satellite images and obtaining the pixel positions of the line towers in the satellite images based on a target detection neural network; step 2: calculating the longitude and latitude of the high-voltage line towers through the pixel positions; step 3: detecting and extracting the shadows of the high-voltage line towers in the satellite images; step 4: calculating the shadow lengths of the high-voltage line towers and inversely calculating the tower heights; step 5: training a graph neural network model, inputting the longitude and latitude of the high-voltage line towers into the trained neural network model to predict the connection probability between the high-voltage line towers; converting the connection probability into a connection prediction result through the group characteristics of the high-voltage line towers; step 6: constructing a high-voltage line database based on the longitude and latitude positions of the high-voltage line towers, the heights of the high-voltage line towers, and the connection relationships of the high-voltage line towers. The present invention can use machine learning methods to replace a large amount of manual operations and on-site measurements to reduce the time for constructing a high-voltage line database.

Description

Technical Field

[0001] The present invention belongs to the technical field of helicopter near - ground warning, and particularly relates to a method for establishing a high - voltage line database for helicopter anti - collision based on machine learning. Background Art

[0002] Due to the characteristics of vertical take - off and landing and hovering in the air, helicopters are widely used in low - altitude missions. During the execution of low - altitude missions, helicopters often crash due to colliding with obstacles. High - voltage lines, due to their small size and being difficult to detect, have become one of the main threats to the safety of helicopter low - altitude flight.

[0003] Currently, the most widely used method for helicopter high - voltage line anti - collision is the cable cutting method. In this method, a cutter is installed on the helicopter fuselage. When the helicopter collides with a high - voltage line, the high - voltage line is cut. However, this method will damage the high - voltage line and the cost is relatively high.

[0004] The Helicopter Terrain Awareness and Warning System (HTAWS) is a helicopter - mounted warning system. HTAWS calculates the proximity of the helicopter to the terrain / obstacles using a terrain / obstacle database and flight parameters. When the helicopter enters the warning area, HTAWS immediately generates an alarm to remind the pilot to take measures to avoid terrain / obstacle collision. Currently, some foreign companies add the high - voltage line database in the flight airspace to the terrain database and use HTAWS to achieve high - voltage line anti - collision warning. This method has achieved good high - voltage line anti - collision effects. Establishing a high - voltage line database based on public information is of great significance for helicopter high - voltage line anti - collision.

[0005] A patent (application patent number: CN202110026211.2, publication date: July 15, 2022, publication number: CN114756635A) discloses a method for creating a high - voltage line tower database using satellite maps. This method manually searches for high - voltage line towers in satellite maps, marks their longitude and latitude positions after finding them, then asks relevant departments about the height of the high - voltage line towers or measures the tower height on - site, and constructs the high - voltage line database by manually connecting the high - voltage line towers. Due to the need for a large amount of manual operation, this method is cumbersome. And some high - voltage line towers are located in remote areas, making it difficult to measure the height on - site. Therefore, this method is only applicable to establishing a small - scale high - voltage line database, unable to establish a large - scale high - voltage line database, and it takes a long time. Summary of the Invention

[0006] In view of the above deficiencies, the present invention provides a method for establishing a high-voltage line database for helicopter anti-collision based on machine learning, which can use machine learning methods to replace a large amount of manual operations and on-site measurements to reduce the time for constructing the high-voltage line database.

[0007] A method for establishing a high-voltage line database based on machine learning includes Step 1: Collect high-voltage line towers in satellite images, make a data set, input the data set into a target detection neural network for training, and input the satellite image containing the high-voltage line tower into the trained target detection neural network to obtain the pixel positions of the high-voltage line tower in the satellite image; Step 2: Calculate the longitude and latitude of the high-voltage line tower through the pixel positions of the high-voltage line tower. Step 3: Detect and extract the shadows of the high-voltage line towers in the satellite image; Step 4: Calculate the shadow length of the high-voltage line tower and invert the tower height through the shadow length; Step 5: Train a graph neural network model, input the longitude and latitude of the high-voltage line tower into the trained neural network model, and predict the connection probability between the high-voltage line towers; Through the group characteristics of the high-voltage line towers, convert the connection probability between the high-voltage line towers into a connection prediction result; Step 6: Construct a high-voltage line database based on the longitude and latitude positions of the high-voltage line towers, the heights of the high-voltage line towers, and the connection relationships between the high-voltage line towers.

[0008] Preferably, Step 1 specifically includes: Step 1.1: Collect pictures of high-voltage line towers in different background environments in satellite images, divide the corner pixels of the high-voltage line towers in the pictures, and complete the production of the high-voltage line tower data set; Step 1.2: Divide the data set into a training set and a validation set, input it into the target detection neural network, set appropriate learning rates and learning step sizes for training, and make the neural network model converge through iterative operations to obtain the weights of the neural network model that can accurately detect high-voltage line towers; Step 1.3: Input the satellite image containing the high-voltage line tower into the trained target detection neural network to obtain the pixel positions of the high-voltage line tower in the satellite image.

[0009] Preferably, in Step 2, according to the satellite image information, convert the pixel positions of the high-voltage line towers into longitude and latitude positions of the high-voltage line towers. The satellite image selects TIFF format image files, which contain image coordinate system and scale information. The longitude and latitude calculation formula of the high-voltage line tower is as follows:

[0010] lon = lon0 + x * e lon + y * R row (1)

[0011] lat = lat0 + y * e lat + x * R col (2)

[0012] Among them, lon and lat are the longitude and latitude of the target pixel point respectively; lon0 and lat0 are the longitude and latitude of the pixel point in the upper left corner of the image respectively; x and y are the pixel positions of the target pixel point; elon and e lat are the longitude and latitude resolutions respectively, that is, the longitude and latitude sizes corresponding to each pixel; R row and R col are the row rotation and column rotation respectively.

[0013] Preferably, in step 3, three thresholds are used for shadow detection: in the RGB color space, the brightness value of the shadow area in the satellite image is lower than that of the non-shadow area, and among the R, G, and B parameters in the shadow color space, the R and G components are smaller and the B component is larger. The ratio of the B component to the R and G components is used as the basis for judging whether it is a shadow area. The color model C3 is used to describe the ratio of the B component to the R and G components, which is expressed as:

[0014]

[0015] The area with a small C3 component is judged as a non-shadow area; at the same time, although the B component is relatively large in the shadow area, due to the low brightness of the shadow area, the absolute value of the B component is not large. Therefore, when a certain area simultaneously satisfies that the C3 component is greater than T1 and the B component is less than T2, it is judged as a shadow area, and T1 and T2 are 0.2 and 90 respectively.

[0016] Preferably, to reduce the misdetection situation where vegetation is detected as a shadow due to low brightness, the G component is selected as a condition, and the pixels with a higher G component are judged as non-shadow areas, that is, when a certain area simultaneously satisfies that the C3 component is greater than T1, the B component is less than T2, and the G component is less than T3, it is judged as a shadow area, and T3 is 60.

[0017] Preferably, step 4 specifically includes step 4.1, inputting the local longitude and latitude and the time when the image is obtained to obtain the solar altitude angle and azimuth angle. The calculation formulas are as follows:

[0018] sinα = sinω × sinδ + cosω × cosδ × cost (4)

[0019] cosβ = (sinα × sinω - sinδ) / (cosα × sinω) (5)

[0020] where α is the solar altitude angle, β is the solar azimuth angle, ω is the latitude, t is the hour angle, and δ is the solar declination;

[0021] Step 4.2, based on the pixel position of the high-voltage tower obtained in step 1, starting from the position of the high-voltage tower, a ray is drawn in the direction of the solar azimuth angle; during the extension of the ray, it is judged whether the pixel where the ray is located is a shadow area. If the foremost point in the ray extension direction is in the shadow area, continue to extend forward. If it is a non-shadow area, calculate the pixel distance from the foremost point in the ray extension direction to the shadow area, and continue to extend along the ray direction;

[0022] Step 4.3: Set a distance threshold. When the pixel distance is greater than the threshold, stop extending the ray and calculate the length of the line segment. The length of the high-voltage tower shadow is the length of the line segment minus the threshold value.

[0023] Step 4.4: Invert the tower height based on the length of the high-voltage tower shadow. The tower height calculation formula is:

[0024] H = S / cotα (6)

[0025] Where S is the shadow length, H is the height of the high-voltage tower, and α is the solar altitude angle.

[0026] Preferably, to reduce the calculation error of the high-voltage tower shadow length, several parallel rays parallel to the ray are drawn. The endpoints of the parallel rays are all on the perpendicular line of the ray and are equidistant; repeat the operations of Step 4.1 - Step 4.3 to obtain the lengths of n line segments, and calculate the average value of n + 1 line segments; use the Z-score outlier detection method to find the standard deviation of the lengths of each line segment, and eliminate the line segments that differ from the average value by more than 2 standard deviations as outliers; after eliminating the outliers, calculate the average value of all line segments again, which is the length of the high-voltage tower shadow.

[0027] Preferably, in Step 5, predict the connection line between high-voltage towers based on the longitude and latitude coordinates of the high-voltage towers. Define the geographical coordinates of the high-voltage towers and the connection line between the high-voltage towers as a graph structure, where the geographical coordinates are the features of the nodes and the connection line is the edge; generalize the high-voltage tower connection line prediction problem to the graph link prediction problem; input the longitude and latitude coordinates of the high-voltage towers into the trained graph neural network, and output the probability of the connection line between the high-voltage towers.

[0028] Preferably, in Step 6, add the high-voltage line database to the existing terrain elevation database: According to the connection relationship between high-voltage towers, calculate the elevation grid passed by the high-voltage tower connection line, and add a high-voltage tower height H to the elevation grid passed by the high-voltage line.

[0029] Preferably, in Step 6, the method for calculating the elevation grid passed by the connection line is: use the points between m high-voltage towers to replace the high-voltage line. First, calculate the longitude and latitude of m points and obtain the corresponding elevation grid, and then add the high-voltage tower height H to the elevation grid. Among them: the longitude and latitude (x, y) of m points = ((x2 - x1) / m * n + x1, (y2 - y1) / m * n + y1), n = 0, 1, 2…, m; the height H = (H1 – H2) / m * n + H2, n = 0, 1, 2…, m; where m is a user-defined value, (x1, y1), (x2, y2) are the longitude and latitude coordinates of high-voltage tower A1 and high-voltage tower A2, and H1, H2 are the heights of high-voltage tower A1 and high-voltage tower A2.

[0030] The beneficial effects of the present invention are as follows:

[0031] (1) The method for establishing a high-voltage line database disclosed by the present invention only uses the public information of satellite images, detects the satellite images by using a target detection neural network, obtains the pixel positions of high-voltage line towers in the satellite images, and calculates the longitude and latitude positions of the high-voltage line towers through the pixel positions, solving the problem that the longitude and latitude of high-voltage line towers need to be manually located. Compared with the existing methods, it does not require a large amount of on-site measurement and does not require a large amount of manual operation, and can quickly construct a high-voltage line database while reducing the cost of constructing the high-voltage line database.

[0032] (2) The method for establishing a high-voltage line database disclosed by the present invention can increase the pilot's perception ability of high-voltage lines compared with the existing helicopter cable cutting method, thereby avoiding collisions with high-voltage lines, not causing damage to high-voltage lines, and protecting the life and property safety of pilots and passengers to the greatest extent.

[0033] (3) Compared with the existing active detection methods for high-voltage lines, it is very difficult to actively detect high-voltage lines because of their small volume. There are cases of missed detection when using the active detection method to detect high-voltage lines, and when the high-voltage lines are detected, the distance between the helicopter and the high-voltage lines is short, and the pilot does not have enough time to escape.

[0034] (4) The method for establishing a high-voltage line database disclosed by the present invention does not require the use of active detection equipment and does not rely on the perception ability of active detection equipment, avoiding the situation of collision with high-voltage lines due to missed detection.

[0035] (5) The method for establishing a high-voltage line database disclosed by the present invention uses the shadow length inversion method to obtain the height of high-voltage line towers, solving the problem that the height of high-voltage line towers cannot be quickly and accurately obtained.

[0036] (6) The method for establishing a high-voltage line database disclosed by the present invention uses a graph neural network to predict the connection of high-voltage line towers, solving the problem that the connection of high-voltage line towers needs to be manually judged. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following will briefly introduce the drawings required for use in the embodiments.

[0038] Figure 1 is the flowchart of the method according to an embodiment of the present invention;

[0039] Figure 2 is the flowchart for obtaining the longitude and latitude of high-voltage line towers according to an embodiment of the present invention;

[0040] Figure 3 is the flowchart for obtaining the height of high-voltage line towers according to an embodiment of the present invention;

[0041] Figure 4 Original satellite image for an embodiment of the present invention;

[0042] Figure 5 Results of shadow detection and high-voltage tower shadow length acquisition for an embodiment of the present invention;

[0043] Figure 6 Flowchart of graph neural network training for an embodiment of the present invention;

[0044] Figure 7 Graph neural network structure diagram for an embodiment of the present invention;

[0045] Figure 8 Structure diagram of graph autoencoder model for an embodiment of the present invention;

[0046] Figure 9 Flowchart of high-voltage tower connection prediction for an embodiment of the present invention;

[0047] Figure 10 Actual connection situation of high-voltage towers in a certain area for an embodiment of the present invention;

[0048] Figure 11 High-voltage tower connection prediction result for an embodiment of the present invention.

[0049] Reference numerals: 101 - target detection module, 102 - longitude and latitude calculation module, 103 - shadow detection module, 104 - tower height calculation module, 105 - connection prediction module, 106 - high-voltage line database establishment module. Detailed implementation manners

[0050] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0051] Embodiment 1

[0052] The present invention addresses the problem of the lack of a high-voltage line database in HTAWS and proposes a method for constructing a high-voltage line database, aiming to further reduce helicopter collisions with high-voltage lines. In view of the problem that it is difficult to obtain a high-voltage line database, the present invention establishes a high-voltage line database based on public information. This method uses machine learning to automatically construct a high-voltage line database, which has the characteristics of short time consumption and low cost compared with manually creating a high-voltage line database.

[0053] Three pieces of information are required to construct the high-voltage line database of the present invention, including the longitude and latitude of high-voltage towers, the height of high-voltage towers, and the connection lines between high-voltage towers. Since high-voltage lines are small in size and difficult to measure, the connection lines between high-voltage towers are used to replace high-voltage lines. This is achieved through the following steps:

[0054] Step 1: Collect high-voltage towers in satellite images and make a dataset. Input the dataset into a target detection neural network for training. Input the satellite image containing high-voltage towers into the trained target detection neural network to obtain the pixel positions of high-voltage towers in the satellite image.

[0055] Step 2: Calculate the longitude and latitude of high-voltage towers based on their pixel positions.

[0056] Step 3: Detect and extract the shadows of high-voltage towers in satellite images.

[0057] Step 4: Calculate the shadow length of high-voltage towers and invert the tower height through the shadow length.

[0058] Step 5: Train a graph neural network model. Input the longitude and latitude of high-voltage towers into the trained neural network model to predict the connection line probability between high-voltage towers. Through the group features of high-voltage towers, convert the connection line probability between high-voltage towers into a connection line prediction result.

[0059] Step 6: Based on the above steps, construct a high-voltage line database with the longitude and latitude positions of high-voltage towers, the heights of high-voltage towers, and the connection line relationships between high-voltage towers.

[0060] The steps of the present invention are described in detail below:

[0061] In Step 1, first make a dataset of high-voltage towers in satellite images. Collect pictures of high-voltage towers in different background environments in satellite images, such as backgrounds of farmland, mountains, hills, plains, etc. Divide the corner pixels of high-voltage towers in these pictures to complete the production of the high-voltage tower dataset. Divide the dataset into a training set and a validation set. Input the dataset into a target detection neural network, set appropriate learning rates and learning step sizes for training, and make the neural network model converge through iterative operations to obtain the weights of the neural network model that can accurately detect high-voltage towers. Finally, input the satellite image containing high-voltage towers into the trained target detection neural network to obtain the pixel positions of high-voltage towers in the satellite image. In this embodiment, the target detection neural network uses the YOLOv5 model. YOLOv5 generates detection boxes in satellite images and calculates the center positions of the detection boxes, which are the pixel positions of high-voltage towers.

[0062] In Step 2, according to the satellite image information, the pixel positions of the high-voltage towers obtained in Step 1 are converted into longitude and latitude positions. The satellite image selects an image file in TIFF (Tag Image File Format, TIFF) format, which contains information such as the image coordinate system and scale. The calculation formula for the longitude and latitude of the high-voltage tower is as follows:

[0063] lon = lon0 + x * e lon + y * R row (1)

[0064] lat = lat0 + y * e lat + x * R col (2)

[0065] Among them, lon and lat are the longitude and latitude of the target pixel point respectively; lon0 and lat0 are the longitude and latitude of the pixel point at the upper left corner of the image respectively; x and y are the pixel positions of the target pixel point; e lon , e lat are the longitude and latitude resolutions respectively, that is, the longitude and latitude sizes corresponding to each pixel point; R row , R col are the row rotation and column rotation respectively.

[0066] In Step 3, the present invention uses three thresholds for shadow detection. In the RGB color space, the brightness value of the shadow area in the satellite image is generally lower than that of the non-shadow area. Especially, the R and G components are smaller, but the B component is larger. The main reason is that in the imaging process with sunlight as the light source, the environment has a stronger scattering effect on the B component. Therefore, the ratio of the B component to the R and G components can be used as the basis for judging whether it is a shadow area. The color model C3 is a value describing the ratio of the B component to the R and G components. The calculation formula of C3 is as follows:

[0067]

[0068] In view of the characteristics that the R and G components of the shadow are smaller and the B component is larger, C3 is selected as the condition, and the area with a small C3 component is judged as a non-shadow area; at the same time, although the B component is relatively large in the shadow area, due to the relatively low brightness of the shadow area, the absolute value of the B component is not large, so the area with a high B component is judged as a non-shadow area; during the shadow detection process, vegetation is often misdetected as a shadow due to its low brightness. In order to reduce misdetection, in view of the characteristic that the G component of the vegetation is relatively high, the G component is selected as the condition, and the pixel with a high G component is judged as a non-shadow area. To sum up, when the C3 component of a certain area is greater than T1, the B component is less than T2, and the G component is less than T3, it is determined as a shadow area.

[0069] In step 4, based on the position of the high-voltage tower in the satellite image obtained in step 1, the location of the shadow of the high-voltage tower is inferred to distinguish the shadows of other objects.

[0070] First, input the local longitude and latitude and the time when the image is obtained to calculate the altitude angle and azimuth angle of the sun. The calculation formulas are as follows:

[0071] sinα = sinω × sinδ + cosω × cosδ × cost (4)

[0072] cosβ = (sinα × sinω - sinδ) / (cosα × sinω) (5)

[0073] Where α is the altitude angle of the sun, β is the azimuth angle of the sun, ω is the latitude, t is the hour angle, and δ is the declination of the sun.

[0074] Then, use the pixel position of the high-voltage tower obtained in step 1. Starting from the position of the high-voltage tower, draw a ray in the direction of the azimuth angle of the sun; during the extension of the ray, judge whether the pixel where the ray is located is a shadow area. If the pixel at the top of the ray is a non-shadow area, calculate the pixel distance from the top of the ray to the shadow area. Set a threshold. When the pixel distance is greater than the threshold, the ray stops extending and calculate the length of the line segment. The length of the shadow of the high-voltage tower is the length of the line segment minus the size of the threshold. To reduce errors, draw 10 rays parallel to this ray and repeat the above operations to calculate the length of the line segment. To reduce errors, use the Z-score outlier detection method to find the standard deviation of the lengths of each line segment. If the length of a certain line segment differs from the mean by 2 standard deviations, it is removed as an outlier. After removing the outliers, calculate the average value, which is the length of the shadow of the high-voltage tower.

[0075] Finally, after obtaining the shadow length, invert the tower height through the shadow length. The tower height calculation formula is:

[0076] H = S / cotα (6)

[0077] Where S is the shadow length, H is the height of the high-voltage tower, and α is the altitude angle of the sun.

[0078] In step 5, predict the connection lines between high-voltage towers based on the longitude and latitude coordinates of the high-voltage towers. Define the geographical coordinates of the high-voltage towers and the connection lines between the high-voltage towers as a graph structure, where the geographical coordinates are the features of the nodes and the connection lines are the edges. Generalize the problem of predicting the connection lines of high-voltage towers to the problem of link prediction of a graph. Input the longitude and latitude coordinates of the high-voltage towers into the trained graph neural network, and the probability of the connection lines between the high-voltage towers is output. In this embodiment, the graph neural network uses a graph autoencoder neural network, such as Figures 6 to 8As shown in the figure, the graph autoencoder is an end-to-end unsupervised learning framework, consisting of an encoder and a decoder. The encoder uses a two-layer graph convolutional neural network (GNC), and the decoder is an inner product layer. In the figure, A represents the connection relationship of high-voltage towers, and X represents the geographical coordinates of high-voltage towers; A and X are input into the encoder to obtain the node feature vector Z. The encoder consists of a graph neural network stacked twice, and the graph convolutional neural network is adopted in the present invention. The decoder is an inner product layer. A* represents the connection probability of high-voltage towers. If the connection relationship between high-voltage towers is not known in advance, A can be an empty array. By default, there is none in the present invention, so actually X is the input, and only the longitude and latitude are input.

[0079] Furthermore, the connection probability between high-voltage towers is converted into a predicted connection using the group features of high-voltage towers. By referring to the relevant national plans for high-voltage towers, the group features of high-voltage towers are as follows:

[0080] (1) The distance between two adjacent high-voltage towers is usually between 50 meters and 1300 meters;

[0081] (2) A high-voltage tower generally only connects to the nearest 4 high-voltage towers;

[0082] (3) A high-voltage tower only connects to two high-voltage towers.

[0083] The present invention creatively uses a graph neural network for predicting the connection of high-voltage towers, turning the problem of predicting the connection of high-voltage towers into a link prediction problem of a graph structure. The graph neural network is used to obtain the connection probability between high-voltage towers (such as 0.3, 0.5), and the existing features of high-voltage towers are used to convert the probability (such as 0.3, 0.5) into the probability (0 or 1) required for constructing a high-voltage database. 1 represents that there is a connection between high-voltage towers, and 0 represents no connection.

[0084] In step 6, the longitude and latitude, height, and predicted connection of high-voltage towers obtained in the above steps are used to construct a high-voltage database. In order to add the high-voltage database to the existing terrain elevation database and implement warnings for high-voltage lines and terrain in HTAWS, the present invention chooses to construct a high-voltage elevation database. First, a high-voltage tower height is added to the terrain elevation where the high-voltage tower is located. The connection between high-voltage towers is used to replace the real high-voltage line, and it is calculated which elevation grids the connection of high-voltage towers passes through. A high-voltage tower height is added to the elevation grids passed by the high-voltage line.

[0085] After detecting the shadows in the satellite images, based on the object detection in Step 1, the position of the high-voltage tower in the satellite images is determined. Rays are drawn from the pixel positions of the high-voltage tower to calculate the shadow length. This step can determine the shadows generated by the high-voltage tower and obtain the length of the shadows of the high-voltage tower. Compared with the prior art in which shadows in satellite images are detected, but in fact it is impossible to determine whether the shadows are generated by high-voltage towers or buildings or other objects and manual judgment is required, it is more intelligent and has higher accuracy.

[0086] The method for establishing a high-voltage line database proposed by the present invention uses machine learning to replace manual search for the longitude and latitude of high-voltage towers in satellite images, improving the efficiency of obtaining the longitude and latitude of high-voltage towers. At the same time, the height of the high-voltage tower is inversely calculated using the shadow length of the high-voltage tower. Compared with field detection of the height of the high-voltage tower, the acquisition speed is increased and the cost is reduced. Machine learning is used to replace manual connection between high-voltage towers, increasing the connection speed and reducing the labor cost at the same time.

[0087] Embodiment 2

[0088] The process of creating a high-voltage line database system based on machine learning in the present invention is as Figure 1 shown. It includes an object detection module 101, a longitude and latitude calculation module 102, a shadow detection module 103, a tower height calculation module 104, a connection prediction module 105, and a high-voltage line database establishment module 106.

[0089] In the object detection module 101, first, a satellite image dataset of high-voltage towers is made. The high-voltage towers in the dataset are divided into labels and the corner pixel coordinates are divided. After that, the dataset is segmented and normalized, and the dataset is divided into a training set and a test set. Secondly, the object detection neural network is trained. The pictures and labels in the training set and the validation set are input into the object detection neural network, and appropriate learning rates and learning step sizes are set for training. Through iterative operations, the neural network model converges, and the weights of the neural network model capable of accurately detecting high-voltage towers are obtained. Finally, the satellite image is input into the trained object detection neural network, and the pixel positions of the high-voltage towers in the satellite image can be output.

[0090] In the longitude and latitude calculation module 102, it is necessary to use the Geospatial Data Abstraction Library to process the satellite image. The format of the satellite image is different from that of ordinary pictures. Its format is TIFF, which contains information such as the image coordinate system and scale. Using the pixel positions of the high-voltage towers and the information such as the satellite image coordinate system and scale, the longitude and latitude of the high-voltage towers are calculated. The calculation formula is as follows:

[0091] lon = lon0 + x * e lon + y * R row (1)

[0092] lat = lat0 + y * e lat + x * R col (2)

[0093] where lon and lat are the longitude and latitude of the target pixel point respectively; lon0 and lat0 are the longitude and latitude of the pixel point at the upper left corner of the image respectively; x and y are the pixel positions of the target pixel point; e lon , e lat are the longitude and latitude resolutions respectively, that is, the longitude and latitude sizes corresponding to each pixel point; R row , R col are the row rotation and column rotation respectively.

[0094] Figure 2 is the flowchart for obtaining the longitude and latitude of the high-voltage tower. The method proposed in this application is used to Figure 3 obtain the longitude and latitude of the high-voltage tower in. The obtained results are shown in Table 1.

[0095] Table 1

[0096]

[0097] Input the obtained longitude and latitude results into the satellite map for searching, and they can all correspond to the manually marked high-voltage towers, proving the effectiveness of the method proposed by the present invention.

[0098] In the shadow detection module 103, the shadows in the satellite image are detected by using the shadow features. The main features of the shadows in the satellite image are: the R and G components are smaller, but the B component is larger. The color model C3 is a value describing the ratio of the B component to the R and G components. When C3 is less than 0.2, it is judged as a non-shadow area. The calculation formula of C3 is as follows:

[0099]

[0100] In the shadow area, although the B component is larger than the R and G components, due to the low brightness of the shadow area and being distinguished from the blue objects, when the B component is greater than 80, it is judged as a non-shadow area.

[0101] Because of the low brightness characteristic of vegetation, it is often misdetected as a shadow. In order to reduce the misdetection, for the characteristic that the G component of vegetation is higher, when the G component is greater than 80, it is judged as a non-shadow area.

[0102] Scan each pixel point in the satellite image. When a pixel point has C3 less than 0.2, B component less than 80, and G component less than 80, it is judged that the pixel point is a shadow area. Figure 3 The satellite image shadow detection results in Figure 4 are as shown.

[0103] In the tower height calculation module 104, the shadow length of the high-voltage tower is extracted and the height of the high-voltage tower is calculated. In view of the problems of complex shadow background in satellite images and difficulty in distinguishing the shadow length of high-voltage towers from that of other objects, this application proposes a method for extracting the shadow length of high-voltage towers. Obtaining the shadow length of a high-voltage tower requires the solar azimuth angle and the solar altitude angle. The solar azimuth angle and the solar altitude angle can be obtained from the longitude and latitude of the satellite image shooting location and the satellite image shooting time. The formulas for calculating the solar azimuth angle and the solar altitude angle are as follows:

[0104] sinα = sinω × sinδ + cosω × cosδ × cost (4)

[0105] cosβ = (sinα × sinω - sinδ) / (cosα × sinω) (5)

[0106] Where α is the solar altitude angle, β is the solar azimuth angle, ω is the latitude, t is the hour angle, and δ is the solar declination.

[0107] The steps for calculating the shadow length of the high-voltage tower are as follows: First, obtain the pixel position of the high-voltage tower in the target detection module 101. Draw a ray starting from the pixel position of the high-voltage tower and in the direction of the solar azimuth angle; further, when the ray is extended forward, judge whether the frontmost point of the ray is in the shadow area. If it is not in the shadow area, calculate the distance between the frontmost point of the ray and the shadow area and judge whether it is greater than 35 pixel distances; if it is greater than 35 pixel distances, stop extending forward and calculate the length of this line segment; the pixel length of the shadow of the high-voltage tower is the line segment length minus 35 pixel distances. The shadow length of the high-voltage tower is the pixel length multiplied by the satellite image resolution. The formula for calculating the height of the high-voltage tower is:

[0108] H = S / cotα (6)

[0109] Where S is the shadow length, H is the height of the high-voltage tower, and α is the solar altitude angle. Figure 4 The red line in the shadow of the high-voltage tower is the shadow length extracted by this method.

[0110] Figure 5 Figure [ID number] is the flowchart for obtaining the height of the high-voltage tower. Table 2 is an example result of comparing the height of the high-voltage tower obtained by using the high-voltage tower height calculation method proposed in the present invention with the actual tower height of the high-voltage tower. As can be seen from Table 2, the accuracy rate of the height of the high-voltage tower obtained by the method of the present invention is 94.5%. The effect of obtaining the height of the high-voltage tower is good and the accuracy is high.

[0111] Table 2

[0112]

[0113] In the connection prediction module 105, the connection between high-voltage towers is predicted, and the connection between high-voltage towers is used to replace the high-voltage line. Define the geographical coordinates of the high-voltage tower and the connection between high-voltage towers as a graph structure, where the geographical coordinates are the features of the nodes and the connection is the edge. Generalize the high-voltage tower connection prediction problem to the link prediction problem of the graph. In the graph structure, the edge relationship of the graph is usually represented by an adjacency matrix. The adjacency matrix is a matrix of size N*N, where N is the number of nodes, and the matrix elements are 0 or 1. 0 indicates that there is no edge between two nodes, and 1 indicates that there is an edge between two nodes. The probability adjacency matrix is a matrix of size N*N. Where N is the number of nodes, and the matrix elements are the probabilities of the existence of an edge between two nodes.

[0114] The training flow chart of high-voltage tower connection prediction is as Figure 6 shown. First, collect the longitude, latitude and connection relationship of high-voltage towers from satellite images for graph neural network training. Use the longitude and latitude of high-voltage towers as the node features of the graph, and the connection relationship as the edges of the graph. Divide the high-voltage tower longitude, latitude and connection relationship dataset into training set, validation set and test set. Input the training set and validation set into the graph neural network for training. Input the longitude and latitude of the high-voltage tower into the graph neural network to obtain the probability adjacency matrix of the high-voltage tower. Compare the probability adjacency matrix of the high-voltage tower with the adjacency matrix and calculate the loss function. Input the loss function into the optimizer to complete one iteration. Adjust the learning rate and learning step size, and through multiple iterative operations, make the graph neural network converge to obtain the model weights of the graph neural network and complete the training of the graph neural network.

[0115] The process of high-voltage tower connection prediction is as Figure 9 shown. Input the longitude and latitude of the high-voltage tower into the trained graph neural network to obtain the probability adjacency matrix of the high-voltage tower. To convert the probability adjacency matrix into an adjacency matrix, the group features of the high-voltage tower need to be used. By referring to the relevant national high-voltage tower plans, the following high-voltage tower group features can be summarized:

[0116] (1) The distance between two adjacent high-voltage towers is usually 50 meters to 1300 meters;

[0117] (2) A high-voltage tower generally only connects to the nearest 4 high-voltage towers;

[0118] (3) A high-voltage tower only connects to two high-voltage towers;

[0119] After obtaining the probability adjacency matrix of the high-voltage tower, judge whether the first and second features are satisfied between two high-voltage towers. If not, change the corresponding probability adjacency matrix element to 0. Based on the third feature, only two elements in each row of the adjacency matrix are 1. Change the two largest elements in each row of the probability adjacency matrix to 1, and change the remaining elements to 0 to obtain the adjacency matrix.

[0120] To detect the accuracy of the predicted connection lines, the evaluation index MCC is introduced, and the calculation formula is as follows:

[0121]

[0122] In the formula, TP represents the number of times a positive sample is determined as a positive sample, TN represents the number of times a negative sample is determined as a negative sample, FP represents the number of times a negative sample is determined as a positive sample, and FN represents the number of times a positive sample is determined as a negative sample.

[0123] Using the high-voltage tower connection prediction method proposed by the present invention, 1000 high-voltage tower connections are tested, and the average MCC value is 0.479.

[0124] Figure 10 is the actual connection of the high-voltage tower in a certain area, Figure 11 is the predicted connection of the high-voltage tower using the method of the present invention. As can be seen from the figure, although there are some incorrect connections, it meets the helicopter warning requirements.

[0125] In the high-voltage line database establishment module 106, a high-voltage line database is constructed by using the obtained high-voltage tower longitude and latitude, high-voltage tower height, and high-voltage tower connection lines. In order to be able to use the high-voltage line database in HTAWS, the present invention increases the high-voltage tower height along the high-voltage line connection on the basis of the terrain elevation database to construct the high-voltage line database. First, calculate the elevation grid where the high-voltage tower is located, and increase the height of the high-voltage tower in the elevation grid where the high-voltage tower is located. According to the connection relationship between high-voltage towers, calculate which elevation grids the connection line between high-voltage towers passes through, and increase the height of a high-voltage tower on the basis of the original terrain elevation for each grid. The steps for calculating the elevation grid passed by the connection line are as follows:

[0126] 1. Obtain the longitude and latitude coordinates (x1, y1), (x2, y2) and heights H1, H2 of the high-voltage towers A1 and A2

[0127] 2. Calculate the longitude and latitude coordinates between m high-voltage towers A1 and A2, that is, the longitude and latitude coordinates of the high-voltage towers. Replace the high-voltage line with the longitude and latitude and height of m points. The calculation formula for the longitude and latitude coordinates of m high-voltage points is:

[0128] 3. (x, y) = ((x2 - x1) / m * n + x1, (y2 - y1) / m * n + y1), n = 0, 1, 2…, m

[0129] 4. Calculate the height between m high-voltage towers A1 and A2, that is, the height of the high-voltage line. The calculation formula is:

[0130] 5. H = (H1 – H2) / m * n + H2, n = 0, 1, 2…, m

[0131] 6. Calculate the elevation grid where the high-voltage line is located based on the longitude and latitude of the high-voltage line, and increase the height H of the high-voltage line on the basis of the original elevation of the elevation grid.

[0132] The elevation database stores elevation data through grids, and each element in the matrix represents the elevation of a region. The m in the above formula is a custom value, which uses the points between m high-voltage towers to replace the high-voltage line. First, calculate the longitude and latitude of these points, then calculate the height, obtain the elevation grid according to the longitude and latitude, and increase the height on the elevation grid.

[0133] The present invention is different from the prior art. By calculating the longitude and latitude between two connected high-voltage towers, for example, calculating the elevation grids where 100 points are located, and increasing the height of a high-voltage tower on the elevation grids where these points are located.

[0134] Finally, it should be noted that the above are only the preferred embodiments of the present invention and are not used to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some of the technical features. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A method for establishing a high-voltage line database based on machine learning, characterized in that: Step 1: Collect high-voltage line towers in satellite images, make a data set, input the data set into a target detection neural network for training, and input the satellite image containing the high-voltage line tower into the trained target detection neural network to obtain the pixel positions of the high-voltage line tower in the satellite image; Step 2: Calculate the longitude and latitude of the high-voltage line tower through the pixel positions of the high-voltage line tower; Step 3: Detect and extract the shadows of high-voltage line towers in satellite images; Step 4: Calculate the shadow length of the high-voltage line tower and invert the tower height through the shadow length; Step 5: Train a graph neural network model, input the longitude and latitude of the high-voltage line tower into the trained neural network model to predict the connection probability between high-voltage line towers; through the group characteristics of the high-voltage line towers, convert the connection probability between high-voltage line towers into a connection prediction result; Step 6: Construct a high-voltage line database based on the longitude and latitude positions of high-voltage line towers, the heights of high-voltage line towers, and the connection relationships between high-voltage line towers.

2. The method according to claim 1, wherein Specifically, Step 1 includes: Step 1.1: Collect pictures of high-voltage line towers in different background environments in satellite images, divide the corner pixels of the high-voltage line towers in the pictures, and complete the production of the high-voltage line tower data set; Step 1.2: Divide the data set into a training set and a validation set, input it into a target detection neural network, set appropriate learning rates and learning step sizes for training, and make the neural network model converge through iterative operations to obtain the weights of the neural network model that can accurately detect high-voltage line towers; Step 1.3: Input the satellite image containing the high-voltage line tower into the trained target detection neural network to obtain the pixel positions of the high-voltage line tower in the satellite image.

3. The method according to claim 1, wherein In Step 2, according to the satellite image information, convert the pixel positions of the high-voltage line towers into longitude and latitude positions of the high-voltage line towers. The satellite image selects a TIFF format image file, which contains image coordinate system and scale information. The longitude and latitude calculation formula of the high-voltage line tower is as follows: lon = lon0 + x * e lon + y * R row (1) lat = lat0 + y * e lat + x * R col (2) Among them, lon and lat are the longitude and latitude of the target pixel point respectively; lon0 and lat0 are the longitude and latitude of the pixel point at the upper left corner of the image; x and y are the pixel positions of the target pixel point; e lon , e lat are the longitude and latitude resolutions respectively, that is, the longitude and latitude sizes corresponding to each pixel point; R row , R col are the row rotation and column rotation respectively.

4. The method according to claim 1, wherein In Step 3, three thresholds are used for shadow detection: in the RGB color space, the brightness value of the shadow area in the satellite image is lower than that of the non-shadow area, and among the R, G, and B parameters in the shadow color space, the R and G components are smaller and the B component is larger. Use the ratio of the B component to the R and G components as the basis for judging whether it is a shadow area, and use the color model C3 to describe the ratio of the B component to the R and G components, which is expressed as: Regions with small C3 components are judged as non-shadow regions; at the same time, although the B component is relatively large in the shadow area, due to the low brightness of the shadow area, the absolute value of the B component is not large. Therefore, when a certain region simultaneously satisfies C3 component is greater than T1 and B component is less than T2, it is determined as a shadow area, where T1 and T2 are set thresholds.

5. The method according to claim 4, characterized in that, To reduce the misdetection situation where vegetation is detected as a shadow due to low brightness, select the G component as a condition, and judge pixels with a higher G component as non-shadow regions, that is, when a certain region simultaneously satisfies that the C3 component is greater than T1, the B component is less than T2, and the G component is less than T3, it is determined as a shadow area, where T3 is a set threshold.

6. The method according to claim 5, characterized in that, Specifically, Step 4 includes: Step 4.1: Input the local longitude and latitude and the time when the image is obtained to obtain the altitude angle and azimuth angle of the sun. The calculation formula is as follows: sinα = sinω × sinδ + cosω × cosδ × cost (4) cosβ = (sinα × sinω - sinδ) / (cosα × sinω) (5) where α is the solar altitude angle, β is the solar azimuth angle, ω is the latitude, t is the hour angle, and δ is the solar declination; Step 4.2: Based on the pixel positions of the high-voltage tower obtained in Step 1, draw a ray starting from the position of the high-voltage tower in the direction of the solar azimuth angle. During the extension of the ray, determine whether the pixel where the ray is located is in the shadow area. If the foremost point in the ray extension direction is in the shadow area, continue to extend forward. If it is in the non-shadow area, calculate the pixel distance from the foremost point in the ray extension direction to the shadow area and continue to extend along the ray direction; Step 4.3: Set a distance threshold. When the pixel distance is greater than the threshold, stop extending the ray and calculate the length of the line segment. The length of the high-voltage tower shadow is the length of the line segment minus the threshold; Step 4.4: Invert the tower height based on the length of the high-voltage tower shadow. The tower height calculation formula is: H = S / cotα (6) where S is the shadow length, H is the height of the high-voltage tower, and α is the solar altitude angle.

7. The method according to claim 6, characterized in that, To reduce the calculation error of the high-voltage tower shadow length, draw several parallel rays parallel to this ray. The endpoints of the parallel rays are all on the perpendicular line of the ray and are equidistant. Repeat the operations in Steps 4.1 - 4.3 to obtain the lengths of n line segments and calculate the mean of the n + 1 line segments. Use the Z-score outlier detection method to find the standard deviation of the lengths of each line segment, and eliminate the line segments that differ from the mean by more than 2 standard deviations as outliers. After eliminating the outliers, calculate the average value of all the line segments again, which is the length of the high-voltage tower shadow.

8. The method according to claim 7, wherein In Step 5, predict the connection lines between high-voltage towers based on the longitude and latitude coordinates of the high-voltage towers. Define the geographical coordinates of the high-voltage towers and the connection lines between them as a graph structure, where the geographical coordinates are the features of the nodes and the connection lines are the edges. Generalize the high-voltage tower connection line prediction problem to the graph link prediction problem. Input the longitude and latitude coordinates of the high-voltage towers into the trained graph neural network, and output the probability of the connection lines between the high-voltage towers.

9. The method according to claim 1, wherein In Step 6, add the high-voltage line database to the existing terrain elevation database: According to the connection relationship between high-voltage towers, calculate the elevation grids passed by the high-voltage tower connection lines, and add a high-voltage tower height H to the elevation grids passed by the high-voltage lines.

10. The method according to any one of claims 1-8, characterized in that, In Step 6, the method for calculating the elevation grids passed by the connection lines is: Replace the high-voltage line with points between m high-voltage towers. First, calculate the longitude and latitude of the m points and obtain the corresponding elevation grids, and then add the high-voltage tower height H to the elevation grids, where: The longitude and latitude (x, y) of the m points = ((x2 - x1) / m * n + x1, (y2 - y1) / m * n + y1), n = 0, 1, 2…, m; The height H = (H1 – H2) / m * n + H2, n = 0, 1, 2…, m; Where m is a user-defined value, (x1, y1) and (x2, y2) are the longitude and latitude coordinates of the high-voltage tower A1 and the high-voltage tower A2, and H1 and H2 are the heights of the high-voltage tower A1 and the high-voltage tower A2.

Images