Wind shear recognition algorithm

Abstract

The invention relates to a wind shear recognition algorithm, which comprises data preprocessing and method recognition, and the data preprocessing comprises the following steps: S1, analyzing original data; s2, segmenting periodic data; s3, multi-channel wind field data; s4, generating a wind shear label; and S5, constructing a data set. According to the wind shear identification algorithm, a wind shear identification method based on a deep neural network is provided, specific radar scanning data is processed to obtain wind shear image data, the identification algorithm based on deep learning is completed, wind shear is identified by using the image data, and high-real-time wind shear identification and early warning requirements are met. Aiming at the laser wind finding radar data, the network model structure is improved, the recognition precision can be improved under the condition that the operand is reduced, automatic recognition of the target wind shear area is achieved through deep learning, conversion of the radar data is achieved through the image generation technology, and wind shear visualization is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of wind shear identification, and in particular to a wind shear identification algorithm. Background Art

[0002] Wind shear is a common natural disaster that poses a serious threat to aircraft safety during takeoff and landing and flight. Wind shear identification provides important data support for the prevention and control of wind shear disasters. In the early days, wind shear data was mainly obtained by traditional meteorological observation methods. This method is time-consuming and labor-intensive and has low real-time performance. Deep learning is an emerging recognition technology in the 20th century. It has good real-time performance, high single-time detection accuracy, and can capture global context, accurate target detection and positioning. It is widely used in many fields. With the continuous advancement of technology, the recognition accuracy of deep learning technology is also getting higher and higher, thus playing a more important role in wind shear recognition.

[0003] Deep learning can quickly identify wind shear information. Compared with conventional identification methods, it has higher recognition accuracy, faster recognition efficiency, and lower economic consumption. It is one of the indispensable identification methods for meteorological disasters such as wind shear. Among them, wind profiling radar has great potential in wind shear, such as the commonly used wind profile radar and laser wind radar. Through these technologies, laser beams are emitted to detect the scattering of aerosols or cloud particles, thereby obtaining wind speed and direction information, and can provide accurate vertical wind field data.

[0004] The identification of wind shear disasters is the top priority of airport meteorological disaster identification and warning. With the in-depth study of wind shear characteristics, some wind shear identification methods have gradually been proposed, which can be mainly divided into two categories: image feature-based recognition algorithms and sensor-based recognition algorithms. Sensor-based recognition algorithms refer to real-time meteorological data obtained by meteorological sensors. Image feature-based recognition algorithms can visualize wind field data, thereby providing strong support for wind shear warning and research. Traditional wind shear recognition methods rely on various artificially designed features to identify and judge wind shear, but are affected by the weather environment, high computational complexity, limited detection range, and interference. , the effect of traditional recognition algorithms is not ideal. In recent years, convolutional neural networks in deep learning have shown strong performance in image classification, target detection, semantic segmentation and other image processing related tasks. Therefore, there are also a small number of researchers using convolutional neural networks to study wind shear recognition, but these studies are still in their infancy, and only a few simple network models have been designed and tried. Using more complex network structures and adopting more targeted designs are effective ways to improve the accuracy of wind shear recognition. For wind shear recognition tasks, how to make the model better pay attention to and learn the essential characteristics of wind shear is the key issue in designing an automatic wind shear recognition model. Summary of the invention

[0005] In view of the shortcomings of the prior art, the present invention provides a wind shear identification algorithm, which has the advantages of using deep learning to realize automatic identification of target wind shear areas, and solves the problems of low accuracy, large workload and poor real-time performance in traditional wind shear deformation identification methods.

[0006] To achieve the above object, the present invention provides the following technical solutions: a wind shear identification algorithm, including data preprocessing and method identification,

[0007] Among them, data preprocessing includes the following steps:

[0008] S1 parses raw data;

[0009] S2 segmentation cycle data;

[0010] S3 multi-channel wind field data;

[0011] S4 wind shear tag generation;

[0012] Construction of S5 dataset;

[0013] Among them, method identification includes the following steps:

[0014] S1 builds a wind shear identification network model

[0015] S2 algorithm improvements.

[0016] Furthermore, the data preprocessing generates wind field visualization data from radar data, including image conversion, geometric transformation and image enhancement. The image conversion converts text wind field data into image wind field data to intuitively observe wind shear characteristics.

[0017] Furthermore, in the data preprocessing S1, the laser wind measuring radar generates horizontal wind field data by transmitting and receiving signals, and obtains visual text data from the generated raw data through coordinate transformation and data transformation. The parsed data includes the azimuth angle, pitch angle, detection distance, spatial coordinates of the point, wind speed, wind direction, and velocity component of the point.

[0018] Furthermore, the text data is observed and analyzed in the data preprocessing S2. Since the time intervals of each text data are inconsistent and the duration of wind shear is also different, it cannot be used for deep learning recognition training. Therefore, the data needs to be segmented. The periodic angle of a radar scan is 17-345 degrees. The segmentation code is written to segment the total data by period according to the angle characteristics, generate data for a radar scan, and finally obtain periodic horizontal wind field data.

[0019] Furthermore, in the data preprocessing S3, the original coordinates are converted into grid coordinates, the maximum and minimum values ​​of the coordinates in the data are calculated, and the appropriate grid interval is determined to ensure that each point can be mapped in the grid, and then the vector information of the wind speed in the periodic data is extracted and mapped in the x, y, and z channels. Finally, the blank spaces in the image need to be filled. In order to avoid the generated pixel values ​​being too extreme and having obvious color differences from the original image, the cubic convolution interpolation method is used to improve the overall quality of the image, so that the generated image is smoother and continuous, and the noise introduced by boundary processing is avoided. The cubic convolution interpolation method requires the target position of the interpolation to determine the corresponding neighborhood in the original image. In the selected neighborhood, a weight is calculated for each pixel point. This weight represents the contribution of the pixel to the interpolation, and a cubic function is used, as shown below:

[0020]

[0021] Where d is the relative distance between the pixel and the interpolation point, a is a user-defined parameter of -0.5, and this function is a cubic interpolation kernel about the relative distance. For the target position, the interpolation is calculated according to the weights of the adjacent pixels and the pixel values. For all target positions that need to be interpolated, the above steps are repeated to calculate the interpolation and finally obtain the wind field plane image.

[0022] Furthermore, in order to meet the training requirements of a large number of data sets for deep learning algorithms, the data preprocessing S4 needs to annotate the generated horizontal wind field map with wind shear. Therefore, the result of lidar detection is annotated by the formula Convert polar coordinates to rectangular coordinates, determine the center position of the annotation box, determine the range of the annotation box by comparing the location where wind shear occurs, and then annotate it in the horizontal wind field map. Finally, generate a label file for the recognition algorithm according to the deep learning label format to complete the labeling.

[0023] In the preprocessing step S4, the core content is the process of wind shear annotation, which is as follows:

[0024] LiDAR detection results: Wind field data obtained using LiDAR, which usually contains information such as wind speed and direction.

[0025] Conversion from polar coordinates to rectangular coordinates: Convert the polar coordinate data (usually angles and distances) obtained by lidar detection into data in a rectangular coordinate system for subsequent processing and analysis.

[0026] Determine the center position of the annotation box: Based on the converted coordinate data, determine the center position of the wind shear occurrence area.

[0027] Determine the range of the annotation box: Determine the boundaries of the annotation box based on the specific location where the wind shear occurs. This range will be used to mark the wind shear area.

[0028] Marking on the wind map: On the horizontal wind map, the wind shear area is visually marked according to the determined marking box range.

[0029] Generate label files: Generate label files for training and identification algorithms in the label format required by deep learning algorithms. These label files will instruct the model which areas are wind shear and which are not.

[0030] Furthermore, in the data preprocessing S5, for deep learning algorithm training, it is necessary to set the categories in the image, set the image categories to wind shear and non-wind shear, and process the image category labels to 0 and 1, where the pixel value within the wind shear instance range is 0 and the non-wind shear is 1. The horizontal wind field image data within a period of time with a higher frequency of wind shear is selected as the wind shear sample database.

[0031] Furthermore, the method identifies that the S1 stroke shear recognition network model is mainly composed of four parts: Input, Backbone, Head, and Detect. Input is a 640*640*3 data input, Backbone is a backbone network composed of CBS, ELAN, and MP-1, Head is mainly composed of CBS, SPPCSPC, E-ELAN, and MP-2, and Detect is three detection modules. First, the image is pre-processed after a series of operations such as data enhancement in the input part, and then sent to the backbone part to extract image features. The extracted features are fused through the Neck module to obtain features of large, medium, and small sizes. The fused features are sent to the detection head, and the results are output after detection.

[0032] Further, the method identification S2 includes an optimization algorithm and a training model, and the optimization algorithm is specifically as follows:

[0033] 1) The CoordConv convolution module is used in the algorithm model head network. CoordConv can introduce additional coordinate information into the input feature map, giving the model stronger location perception ability and improving the model mAP index;

[0034] 2) The attention mechanism adopts the CBAM dual-channel attention module to achieve attention adjustment in both channel and space dimensions, effectively emphasize the feature channels with higher wind shear, suppress redundant and useless features, accurately extract the key features of wind shear images, and improve the mAP index of the recognition algorithm on the wind shear dataset;

[0035] 3) Using the SPPFCSPC module as the pooling layer, the recognition speed is improved by 18% while keeping the receptive field unchanged, saving computing costs;

[0036] 4) The recognition algorithm adopts the Fre lu activation function as the Fre lu activation function to enhance the nonlinear expression ability of the network and improve the monitoring accuracy without significantly increasing the calculation amount of the model.

[0037] Furthermore, the training model uses the labeled wind shear sample data set to train the wind shear recognition model. Finally, the wind shear data set is divided into a training set and a validation set in a ratio of 8:2. The neural network is iteratively trained to obtain the optimal wind shear area recognition model.

[0038] Compared with the prior art, the technical solution of this application has the following beneficial effects:

[0039] The wind shear recognition algorithm provides a wind shear deformation recognition method based on a deep neural network. By processing specific radar scanning data, wind shear image data is obtained, and a recognition algorithm based on deep learning is completed. Wind shear is recognized using image data to achieve high real-time wind shear recognition and early warning requirements. The network model structure is improved for the laser wind measurement radar data in the present invention, which can improve the recognition accuracy while reducing the amount of calculation. Deep learning is used to realize automatic recognition of the target wind shear area, and image generation technology is used to realize the conversion of radar data to realize wind shear visualization. The algorithm solves the problems of low real-time wind shear recognition and low recognition accuracy, and is suitable for wind shear recognition research of other radar equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 It is a schematic diagram of the data preprocessing process of the present invention;

[0041] Figure 2 This is a schematic diagram of the total data visualization image of the present invention;

[0042] Figure 3 It is a schematic diagram of the product interpolation result of the present invention;

[0043] Figure 4 It is a schematic diagram of a horizontal wind field of the present invention;

[0044] Figure 5 This is a schematic diagram of the structure of the wind shear identification model of the present invention;

[0045] Figure 6 This is a schematic diagram showing the recognition results of the present invention. DETAILED DESCRIPTION

[0046] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0047] See also Figure 1-6 , a wind shear identification algorithm in this embodiment includes data preprocessing and method identification,

[0048] Among them, data preprocessing includes the following steps:

[0049] S1 parses raw data;

[0050] S2 segmentation cycle data;

[0051] S3 multi-channel wind field data;

[0052] S4 wind shear tag generation;

[0053] Construction of S5 dataset;

[0054] Among them, method identification includes the following steps:

[0055] S1 builds a wind shear identification network model

[0056] S2 algorithm improvements.

[0057] It should be noted that, based on the characteristics of the data set, the algorithm selects the coordconv convolutional network as the backbone network, supplemented by the CBAM attention mechanism to strengthen the extraction of image features, uses the SPPFCSPC pyramid pooling module to improve the algorithm recognition speed, and uses the FReLU function as the activation function to achieve fast and accurate meteorological disaster identification.

[0058] Among them, data preprocessing converts radar data into wind field visualization data, including image conversion, geometric transformation and image enhancement. Image conversion converts text wind field data into image wind field data to intuitively observe wind shear characteristics.

[0059] In general, image recognition data preprocessing is an important step, which can improve the accuracy and efficiency of image recognition and provide a better data foundation for subsequent operations.

[0060] Among them, in data preprocessing S1, the laser wind measurement radar generates horizontal wind field data by transmitting and receiving signals, and the generated raw data is converted into visual text data through coordinate conversion and data conversion. The parsed data includes the azimuth, pitch angle, detection distance, spatial coordinates of the point, wind speed, wind direction, and velocity component of the point.

[0061] Table 1 Text data description

[0062]

[0063]

[0064] Among them, in data preprocessing S2, text data is observed and analyzed. Since the time intervals of each text data are inconsistent and the duration of wind shear is also different, it cannot be used for deep learning recognition training. Therefore, the data needs to be segmented. The periodic angle of a radar scan is 17-345 degrees. The segmentation code is written to segment the total data by period according to the angle characteristics, generate data for a radar scan, and finally obtain periodic horizontal wind field data.

[0065] Among them, in the data preprocessing S3, the original coordinates are converted into grid coordinates, the maximum and minimum values ​​of the coordinates in the data are calculated, and the appropriate grid interval is determined to ensure that each point can be mapped in the grid, and then the vector information of the wind speed in the periodic data is extracted and mapped in the x, y, and z channels. Finally, the blank spaces in the image need to be filled. In order to avoid the generated pixel values ​​being too extreme and the color difference from the original image being obvious, the cubic convolution interpolation method is used to improve the overall quality of the image, making the generated image smoother and continuous, and avoiding the noise introduced by boundary processing. The cubic convolution interpolation method requires the target position of the interpolation to determine the corresponding neighborhood in the original image. In the selected neighborhood, a weight is calculated for each pixel point. This weight represents the contribution of the pixel to the interpolation. A cubic function is used as shown below:

[0066]

[0067] Where d is the relative distance between the pixel and the interpolation point, a is a user-defined parameter of -0.5, and this function is a cubic interpolation kernel about the relative distance. For the target position, the interpolation is calculated according to the weights of the adjacent pixels and the pixel values. For all target positions that need to be interpolated, the above steps are repeated to calculate the interpolation and finally obtain the wind field plane image.

[0068] In the data preprocessing S4, in order to meet the training requirements of a large number of data sets for deep learning algorithms, it is necessary to annotate the generated horizontal wind field map with wind shear. Therefore, the results of lidar detection are annotated by the formula Convert polar coordinates to rectangular coordinates, determine the center position of the annotation box, determine the range of the annotation box by comparing the location where wind shear occurs, and then annotate it in the horizontal wind field map. Finally, generate a label file for the recognition algorithm according to the deep learning label format to complete the labeling.

[0069] Among them, in data preprocessing S5, for deep learning algorithm training, it is necessary to set the category in the image, set the image category to wind shear and non-wind shear, and process the image category labels to 0 and 1, where the pixel value within the wind shear instance range is 0 and the non-wind shear is 1. The horizontal wind field image data within a period of time with a high frequency of wind shear is selected as the wind shear sample database.

[0070] Among them, the method identifies that the S1 stroke shear recognition network model is mainly composed of four parts: Input, Backbone, Head, and Detect. Input is the data input of 640*640*3, Backbone is the backbone network composed of CBS, ELAN, and MP-1, Head is mainly composed of CBS, SPPCSPC, E-ELAN, and MP-2, and Detect is three detection modules. First, the image is pre-processed after a series of operations such as data enhancement in the input part, and then sent to the backbone part to extract image features. The extracted features are fused through the Neck module to obtain features of large, medium, and small sizes. The fused features are sent to the detection head, and the results are output after detection.

[0071] Among them, method identification S2 includes an optimization algorithm and a training model. The optimization algorithm is as follows:

[0072] 1) The CoordConv convolution module is used in the algorithm model head network. CoordConv can introduce additional coordinate information into the input feature map, giving the model stronger location perception ability and improving the model mAP index;

[0073] 2) The attention mechanism adopts the CBAM dual-channel attention module to achieve attention adjustment in both channel and space dimensions, effectively emphasize the feature channels with higher wind shear, suppress redundant and useless features, accurately extract the key features of wind shear images, and improve the mAP index of the recognition algorithm on the wind shear dataset;

[0074] 3) Using the SPPFCSPC module as the pooling layer, the recognition speed is improved by 18% while keeping the receptive field unchanged, saving computing costs;

[0075] 4) The recognition algorithm adopts the Fre lu activation function as the Fre lu activation function to enhance the nonlinear expression ability of the network and improve the monitoring accuracy without significantly increasing the calculation amount of the model.

[0076] Among them, the training model uses the labeled wind shear sample data set to train the wind shear recognition model. Finally, the wind shear data set is divided into a training set and a validation set in a ratio of 8:2. The neural network is iteratively trained to obtain the optimal wind shear area recognition model.

[0077] The beneficial effects of the above embodiments are: providing a wind shear deformation recognition method based on a deep neural network, obtaining wind shear image data by processing specific radar scanning data, completing a recognition algorithm based on deep learning, and using image data to recognize wind shear, thereby achieving high real-time wind shear recognition and early warning requirements. For the laser wind measurement radar data in the present invention, the network model structure is improved, which can improve the recognition accuracy while reducing the amount of calculation, and use deep learning to realize automatic recognition of the target wind shear area, use image generation technology to realize the conversion of radar data, and realize wind shear visualization, which solves the problems of low real-time wind shear recognition and low recognition accuracy, and is suitable for wind shear recognition research of other radar equipment.

[0078] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0079] While the embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that many changes, modifications, substitutions and variations can be made to the embodiments without departing from the principles and spirit of the invention.

Claims

1. A wind shear identification algorithm, characterized in that: Including data preprocessing and method identification, Among them, data preprocessing includes the following steps: S1 parses raw data; S2 segmentation cycle data; S3 multi-channel wind field data; S4 wind shear tag generation; Construction of S5 dataset; Among them, method identification includes the following steps: S1 builds a wind shear identification network model S2 algorithm improvements.

2. A wind shear identification algorithm according to claim 1, characterized in that: The data preprocessing generates wind field visualization data from radar data, including image conversion, geometric transformation and image enhancement. The image conversion converts text wind field data into image wind field data to visually observe wind shear characteristics.

3. A wind shear identification algorithm according to claim 1, characterized in that: In the data preprocessing S1, the laser wind measuring radar generates horizontal wind field data by transmitting and receiving signals, and obtains visual text data from the generated raw data through coordinate conversion and data conversion. The parsed data includes the azimuth angle, pitch angle, detection distance, spatial coordinates of the point, wind speed, wind direction, and velocity component of the point.

4. A wind shear identification algorithm according to claim 1, characterized in that: In the data preprocessing S2, the text data is observed and analyzed. Since the time intervals of each text data are inconsistent and the duration of wind shear is also different, it cannot be used for deep learning recognition training. Therefore, the data needs to be segmented. The periodic angle of a radar scan is 17-345 degrees. The segmentation code is written to segment the total data by period according to the angle characteristics, generate data for a radar scan, and finally obtain periodic horizontal wind field data.

5. A wind shear identification algorithm according to claim 1, characterized in that: In the data preprocessing S3, the original coordinates are converted into grid coordinates, the maximum and minimum values ​​of the coordinates in the data are calculated, and the appropriate grid interval is determined to ensure that each point can be mapped in the grid, and then the vector information of the wind speed in the periodic data is extracted and mapped in the x, y, and z channels. Finally, the blank spaces in the image need to be filled. In order to avoid the generated pixel values ​​being too extreme and the color difference from the original image being obvious, the cubic convolution interpolation method is used to improve the overall quality of the image, so that the generated image is smoother and continuous, and the noise introduced by boundary processing is avoided. The cubic convolution interpolation method requires the target position of the interpolation to determine the corresponding neighborhood in the original image. In the selected neighborhood, a weight is calculated for each pixel point. This weight represents the contribution of the pixel to the interpolation. A cubic function is used as shown below: Where d is the relative distance between the pixel and the interpolation point, a is a user-defined parameter of -0.5, and this function is a cubic interpolation kernel about the relative distance. For the target position, the interpolation is calculated according to the weights of the adjacent pixels and the pixel values. For all target positions that need to be interpolated, the above steps are repeated to calculate the interpolation and finally obtain the wind field plane image.

6. A wind shear identification algorithm according to claim 1, characterized in that: In the data preprocessing S4, in order to meet the training requirements of a large number of data sets for deep learning algorithms, it is necessary to annotate the generated horizontal wind field map with wind shear. Therefore, the results of lidar detection are annotated by the formula Convert polar coordinates to rectangular coordinates, determine the center position of the annotation box, determine the range of the annotation box by comparing the location where wind shear occurs, and then annotate it in the horizontal wind field map. Finally, generate a label file for the recognition algorithm according to the deep learning label format to complete the labeling.

7. A wind shear identification algorithm according to claim 1, characterized in that: In the data preprocessing S5, for deep learning algorithm training, it is necessary to set the category in the image, set the image category to wind shear and non-wind shear, and process the image category labels to 0 and 1, where the pixel value within the wind shear instance range is 0 and the non-wind shear is 1. The horizontal wind field image data within a period of time with a high frequency of wind shear is selected as the wind shear sample database.

8. A wind shear identification algorithm according to claim 1, characterized in that: The method identifies S1 stroke shear recognition network model mainly composed of four parts: Input, Backbone, Head, and Detect. The Input is a 640*640*3 data input. The Backbone is the backbone network composed of CBS, ELAN, and MP-1. The Head is mainly composed of CBS, SPPCSPC, E-ELAN, and MP-2. Detect is three detection modules. First, the image is pre-processed through a series of operations such as data enhancement in the input part, and then sent to the backbone part to extract image features. The extracted features are fused through the Neck module to obtain features of three sizes: large, medium, and small. The fused features are sent to the detection head, and the results are output after detection.

9. A wind shear identification algorithm according to claim 1, characterized in that: The method identification S2 includes an optimization algorithm and a training model, and the optimization algorithm is specifically as follows: 1) The CoordConv convolution module is used in the algorithm model head network. CoordConv can introduce additional coordinate information into the input feature map, giving the model stronger location perception ability and improving the model mAP index; 2) The attention mechanism adopts the CBAM dual-channel attention module to achieve attention adjustment in both channel and space dimensions, effectively emphasize the feature channels with higher wind shear, suppress redundant and useless features, accurately extract the key features of wind shear images, and improve the mAP index of the recognition algorithm on the wind shear dataset; 3) Using the SPPFCSPC module as the pooling layer, the recognition speed is improved by 18% while keeping the receptive field unchanged, saving computing costs; 4) The recognition algorithm uses the Frelu function as the Frelu activation function to enhance the nonlinear expression ability of the network and improve the monitoring accuracy without significantly increasing the calculation amount of the model.

10. A wind shear identification algorithm according to claim 9, characterized in that: The training model uses a labeled wind shear sample data set to train a wind shear recognition model. Finally, the wind shear data set is divided into a training set and a validation set in a ratio of 8:

2. The neural network is iteratively trained to obtain an optimal wind shear area recognition model.

Images