Gust front wind shear identification method based on artificial intelligence

Abstract

The invention provides a gust front wind shear identification method based on artificial intelligence, and the method comprises the steps: carrying out the noise filtering, missing value supplementary measurement, data smoothing, wind shear value calculation and sample screening extraction of collected radial speed data, and obtaining a gust front wind shear sample; performing coordinate system conversion, sample set division and data annotation on the basis of gust and front wind shear samples to obtain an expanded data set; designing and training a Mask R-CNN model architecture to obtain a gust and front wind shear identification model; and inputting the expanded data set into the gust and front wind shear identification model to carry out gust and front wind shear detection. According to the method, dependence on reflectivity factor data can be reduced, an identification model is constructed based on gust front radial speed data, gust front wind shear can be accurately identified, pixel-level segmentation and positioning of a wind shear area can be realized, and identification efficiency is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of gust front wind shear identification, in particular to a gust front wind shear identification method based on artificial intelligence. BACKGROUND

[0002] Gust front is a kind of mesoscale disastrous weather phenomenon, which is currently mainly monitored by Doppler weather radar. Gust front wind shear appears as a band-shaped wind convergence or wind shear area in the radar velocity field. The existing gust front wind shear identification algorithm mainly includes two categories: one is based on the narrowband echo feature in the radar reflectivity factor image, such as the bidirectional gradient method and the double-template local binary segmentation; the other is based on the radial velocity feature of gust front, such as the detection algorithm proposed by Uyeda, which realizes wind shear identification by fusing the radial convergence and azimuth shear double features, combining threshold screening and feature classification, but the false positive rate of this method is high in complex environment, and the robustness is insufficient. The gust front wind shear algorithm developed by Lincoln Laboratory based on fuzzy logic algorithm and image processing technology has significantly improved the identification accuracy and realized business application, but it has higher requirements for radar data quality, and there are certain limitations in the scope of application.

[0003] Most of the domestic and foreign schemes do not make full use of the "radial velocity", or can only realize rough qualitative detection of the shear area, and the segmentation and positioning are not accurate enough, the process is complicated, and it is difficult to realize business automation. In recent years, deep learning technology has been gradually applied to intelligent identification of disastrous weather. When identifying gust front wind shear by using deep learning technology, there are certain limitations: on the one hand, most methods mainly use reflectivity factor data of gust front, and the utilization rate of radial velocity data closely related to gust front wind shear is low; on the other hand, some algorithms can only realize rough positioning of the wind shear area, and cannot complete pixel-level accurate segmentation, and the process is relatively complex, which has many problems in actual business application. SUMMARY

[0004] In order to overcome the deficiencies of the prior art, the purpose of the present application is to provide a gust front wind shear identification method based on artificial intelligence, which can reduce the dependence on reflectivity factor data, focus on constructing an identification model based on gust front radial velocity data, accurately identify gust front wind shear, realize pixel-level segmentation and positioning of the wind shear area, and improve the identification efficiency.

[0005] To achieve the above purpose, the present application provides the following scheme: a gust front wind shear identification method based on artificial intelligence, comprising:

[0006] Radial velocity data is collected by using a new generation of Doppler weather radar, and the radial velocity data is subjected to noise filtering, missing value filling, data smoothing, wind shear value calculation and sample screening and extraction to obtain gust front wind shear samples;

[0007] Based on the gust front wind shear sample, coordinate system conversion, sample set division and data labeling are performed to obtain an expanded data set;

[0008] A Mask R-CNN model architecture including a feature extraction module, a region candidate network module, a RoI Align module and a multi-task output module is designed, and the Mask R-CNN model architecture is trained to obtain a gust front wind shear identification model;

[0009] The expanded data set is input into the gust front wind shear identification model for gust front wind shear detection.

[0010] Optionally, radial velocity data is collected by using a new generation Doppler weather radar, and the radial velocity data is subjected to noise filtering, missing value filling, data smoothing, wind shear value calculation and sample screening and extraction to obtain a gust front wind shear sample, including:

[0011] The radial velocity data of a 0.5° elevation layer 1-2 km above the ground is selected by using PPI under VCP 21 mode for scanning by using a new generation Doppler weather radar;

[0012] Based on the radial velocity data, noise filtering and missing value filling are performed by using a sliding window frequency method, and the key variable parameters of the sliding window are tested by using a control variable method to screen an optimal parameter combination suitable for the radial velocity data;

[0013] The radial velocity data is subjected to smoothing processing by using a median filtering method with a filtering window of 3x3, when all data points in the filtering window are valid values, the average value of 8 data points in the filtering window is selected as the center point value, until the full image filtering processing is completed;

[0014] Based on the radial velocity data, the velocity shear band is linearly fitted by using a least square method to calculate radial shear and azimuthal shear, and the radial shear and the azimuthal shear are combined to obtain a combined shear;

[0015] According to the radar physical parameter characteristics and the combined shear threshold, strong wind shear class samples and weak wind shear class samples in the radial velocity data are screened, and feature regions of the strong wind shear class samples and the weak wind shear class samples are extracted to obtain gust front wind shear samples including two types of samples.

[0016] Optionally, the calculation expression of the radial shear is:

[0017] ;

[0018] wherein, is the radial velocity, is the distance to the radar center, is the coefficient;

[0019] The calculation expression of the azimuth shear is:

[0020] ;

[0021] wherein, is the azimuth angle;

[0022] The calculation expression of the combined shear is:

[0023] .

[0024] Optionally, based on the gust front wind shear sample, coordinate system conversion, sample set division and data labeling are performed to obtain an expanded data set, including:

[0025] The polar coordinate system of the gust front wind shear sample is converted into a rectangular coordinate system, and then the gust front wind shear sample is mapped into a 200x200km, 588x588 resolution image to construct a strong wind shear data set and a weak wind shear data set;

[0026] 80% of the data in the strong wind shear data set and the weak wind shear data set are extracted respectively to divide the training set and the validation set, and the remaining 20% of the data are used as the prediction set;

[0027] Using the LabelMe tool, the gust front wind shear features in the training set and the validation set are marked with dots to draw a closed polygon around the outer edge of the convergence line to obtain a labeled wind shear area, and based on the pixel coordinates of the closed curve, a JSON label file containing wind shear position information is generated to obtain the expanded data set.

[0028] Optionally, a Mask R-CNN model architecture is designed, which includes a feature extraction module, a region candidate network module, a RoI Align module and a multi-task output module, including:

[0029] The residual block and the short circuit mechanism are introduced into the ResNet50 to obtain a residual structure, through which an input item directly connected is introduced into each convolution calculation output layer of the ResNet50 to output a multi-level deep feature map, and then the low-layer high-resolution features and the high-layer semantic features in the multi-level deep feature map are fused by using the feature pyramid network to obtain a multi-scale feature map, thereby completing the design of the feature extraction module;

[0030] For each pixel point in the multi-scale feature map, 9 candidate boxes combined randomly by 3 different aspect ratios and 3 different scales are generated, and a Softmax classification is used to determine whether each candidate box contains a gust front wind shear target, so as to correct the position and size of the candidate box and obtain an initial candidate box. Then, a non-maximum suppression algorithm with an IoU threshold of 0.7 is used to filter redundant boxes of the initial candidate box to obtain a candidate region, and the design of the region candidate network module is completed.

[0031] The candidate region is mapped to the multi-scale feature map by using a bilinear interpolation method, and the multi-scale feature map is pooled to obtain a 7*7 distortion-free alignment feature map, and the design of the RoI Align module is completed.

[0032] A multi-task output module including a classification branch, a bounding box regression branch and a mask branch is designed, and the feature extraction module, the region candidate network module, the RoI Align module and the multi-task output module are combined to obtain a Mask R-CNN model architecture.

[0033] Optionally, the classification branch is used to output a binary classification result of gust front wind shear and non-gust front wind shear through a fully connected layer, and a cross-entropy loss function is used to optimize the binary classification result.

[0034] The bounding box regression branch is used to output the coordinate parameters of the minimum bounding matrix of the gust front wind shear, and to minimize the deviation between the predicted box and the real box.

[0035] The mask branch is used to generate a 28*28 pixel binary mask for each detected gust front wind shear through a fully convolutional network layer, and a binary cross-entropy loss function is used to optimize the pixel-level segmentation accuracy.

[0036] Optionally, the Mask R-CNN model architecture is trained to obtain a gust front wind shear identification model, including:

[0037] The strong wind shear class samples and the weak wind shear class samples in the training set are mixed at a ratio of 1:1, and 4 samples are used as a batch for 20 rounds of iterative training. When the mAP_50 index of the validation set is greater than or equal to 80%, the first stage training is completed, and the second stage training is entered. Each batch includes 2 strong wind shear class samples and 2 weak wind shear class samples.

[0038] After the first stage training is completed, the proportion of the weak wind shear class samples is increased to 1:3, the learning rate is reduced, and different training periods are compared horizontally. The weight parameters during all training periods are saved to obtain model parameters, and the second stage training is completed.

[0039] After the second stage training is completed, based on the model parameters, the average precision mean and total loss value are evaluated on the validation set to obtain a training evaluation result, the training optimization of the Mask R-CNN model architecture is completed, and a gust front wind shear identification model is obtained.

[0040] An evaluation index is calculated by using a confusion matrix to evaluate the detection result of the gust front wind shear identification model, and the model optimization is completed.

[0041] Optionally, the confusion matrix includes true positives, false positives, false negatives and true negatives, and the evaluation index includes hit rate, false detection rate, missed detection rate and critical success index.

[0042] The present application provides a gust front wind shear identification method based on artificial intelligence, which has the following technical effects:

[0043] 1) Data acquisition and preprocessing: the complex noise of radar data itself can be eliminated, adaptive filtering and supplementary measurement can be realized by using a sliding window statistical method, and more rich multi-source weather data fusion can be introduced, such as multi-layer elevation angle and more radar physical quantities. The dependence on radar reflectivity factor data is significantly reduced, the radial velocity data closely related to the gust front wind shear is mainly used, and the precision and reliability of the gust front wind shear identification are improved.

[0044] 2) Data-label set construction: polar-rectangular coordinate conversion and high-resolution unification provide an excellent data basis for the adaptation of the back-end deep learning algorithm. Type-based training is beneficial to balance the model generalization ability.

[0045] 3) Model training: by using an advanced target detection and segmentation integrated framework, the deep network gradient flow can be optimized, the hyperparameters can be dynamically adjusted, the strong-weak wind shear sample ratio can be dynamically switched, the small sample-weak feature learning ability of the model can be improved, and the overall recognition precision is improved.

[0046] 4) Model evaluation and result output: by using closed curve segmentation and pixel-level mask prediction, the positioning precision, evaluation dimension, prediction stability and overall recognition precision are guaranteed. The test set of the present application can be expanded to cover more extreme and complex weather conditions.

[0047] The technical solutions of the present application will be further described in detail below with reference to the drawings and embodiments. BRIEF DESCRIPTION OF DRAWINGS

[0048] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0049] Figure 1 A schematic diagram of a method flow chart provided in an embodiment of the present invention;

[0050] Figure 2 A flowchart of the method provided by an embodiment of the present invention;

[0051] Figure 3 Flowchart of the automatic identification algorithm of gust front wind shear based on Mask R-CNN provided by an embodiment of the present invention;

[0052] Figure 4 A schematic diagram of the residual block structure provided by an embodiment of the present invention;

[0053] Figure 5 A graph showing how the average precision and total loss change with the number of training rounds provided by the embodiment of the present invention;

[0054] Figure 6 This is a diagram of radial velocity identification results at five typical moments during two gust front wind shear processes provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0055] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. 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 making creative efforts are within the scope of protection of the present invention.

[0056] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0057] like Figure 1 As shown, the present invention provides a gust front wind shear identification method based on artificial intelligence, comprising:

[0058] 1. Using a new generation Doppler weather radar to collect radial velocity data, the radial velocity data is subjected to noise filtering, missing value supplementation, data smoothing, wind shear value calculation, and sample screening and extraction to obtain gust front wind shear samples; specifically, the following steps are involved:

[0059] 1.1 Using a new generation Doppler weather radar, we used PPI scanning in VCP 21 mode. This mode acquires echo intensity, radial velocity, and other data within the radar's detection range through horizontal circular scanning. Because gust front wind shear typically occurs at an altitude of 1-2 km above the ground, radial velocity data at the 0.5° elevation angle were preprocessed.

[0060] 1.2 Noise filtering and missing value supplementation

[0061] Based on the radial velocity data, the sliding window frequency method is used to filter out noise and supplement missing values, and the key variable parameters of the sliding window are tested by the control variable method to screen the optimal parameter combination suitable for the radial velocity data.

[0062] The principle of the "sliding window frequency method" is to set a small sliding window of M×N distance bins, traverse the radial velocity data of the 0.5° elevation layer in the PPI mode through the small window, regard the center of the small window as the distance bin to be processed, and the processing result is determined by the numerical statistics of all distance bins in the small window. The specific method is:

[0063] (1) Divide the radar data into n intervals of equal distance in ascending order, set the interval between each interval to Δd (dB), and set the value at the center of the window to Z ij , where i and j represent the range library and radial azimuth numbers respectively, and Z ij ±λΔd (λ is a constant) is set as the n+1th interval.

[0064] (2) Count the frequency of each value in the sliding window in different intervals (F1, F2…F n , F n+1 ), let F max is the maximum frequency, Z m It is the median of the interval corresponding to the maximum frequency value Fmax.

[0065] (3) Use formula (1) to determine the value Z of the current window center ij , and according to different statistical results, the preprocessed result Zc is obtained.

[0066] Zc= (1)

[0067] In the formula, k1 is the threshold for noise filtering, k2 is the threshold for supplementary measurement of missing values, and “ND” means no valid data. Formula (1) uses the principle of continuity of spatial variation of precipitation echoes. The physical meaning of different statistical results can be understood as follows: if the value Z at the center of the window is ij is valid data, and when F n+1≤k1, it means that there is only a small amount of data near the center of the window, so it is judged as noise and should be removed; when F n+1 >k1, it means that there is more data near the center of the window, which should be retained. For missing values, when the value Z ij of the window center point is ND, if F max ≥k2, it means that more data in the same interval is detected around the window center, so the middle value Z m of the frequency maximum interval can be assigned to it for value completion; on the contrary, if F max <k2, it is considered that there is less valid data around the window center, so Z ij does not need to be completed.

[0068] The setting of values such as M, N, n, λΔd, k1 and k2 in formula (1) has a certain influence on the preprocessing result, so control variable test needs to be carried out to compare the effects of noise filtering and missing value completion under different parameter conditions, and the optimal parameter configuration is selected for data preprocessing. After the test in this paper, the optimal parameter configuration is obtained as follows: M×N=3×3, n=15, λ=2, Δd=4, k1=3, k2=4.

[0069] 1.3 Data smoothing

[0070] In order to eliminate the pulsatility of data, the median filtering method with a filtering window of 3×3 is used to smooth the radial velocity data. When all data points in the filtering window are valid values, the average value of 8 data points in the filtering window is selected as the center point value, and the filtering process is completed until the whole graph is filtered.

[0071] 1.4 Wind shear value calculation based on least square method

[0072] In order to represent the geometric shape of the radial velocity convergence line / shear line of the gust front, based on the radial velocity data, the least square method is used to linearly fit the velocity shear band, and the radial shear is defined as the change of radar radial velocity along the radial distance. A "fitting window" is selected to calculate the radial shear, which needs to contain n data points in the same radial, i.e. (v1, r1), (v i , r i )…(v n , r n )(v i represents the radial velocity, and r i represents the distance to the radar center). The regression equation is:

[0073] (2)

[0074] The relative error between the estimated value and the actual measured value is:

[0075] (3)

[0076] In order to obtain the best fitting straight line, the above calculated relative error square sum needs to be minimized, that is, the value of D in the following formula is minimized:

[0077] (4)

[0078] Obtain:

[0079] (5)

[0080] (6)

[0081] Solve:

[0082] (7)

[0083] Equation (7) is the radial shear obtained.

[0084] The azimuthal shear is defined as the radial velocity on the same distance circle from the radar, starting from 0° azimuth and rotating clockwise to 360°, with the change of azimuth angle, and the calculation process of the radial shear is the same, and the azimuthal shear is obtained:

[0085] (8)

[0086] The radial shear and the azimuthal shear are combined to obtain the combined shear:

[0087] (9)

[0088] 1.5 Determination and extraction of wind shear line

[0089] According to the characteristics of radar physical parameters and the threshold of combined shear, the strong wind shear samples and the weak wind shear samples in the radial velocity data are screened, and the feature regions of the strong wind shear samples and the weak wind shear samples are extracted, to obtain the gust front wind shear samples including two types of samples. Specifically, it includes:

[0090] In order to more accurately identify the gust front wind shear, the application screens the gust front samples according to the radar physical parameter characteristics of the gust front. These characteristics include: the "narrow band echo" feature presented in the reflectivity factor map; the convergence line or shear line appearing in the radial velocity map; the significant high spectral width value in the spectral width map; the regular moving feature; and the existence of the parent thunderstorm triggering the gust front near it, etc. Based on the above characteristics, the obtained radar data is screened, and the samples meeting the conditions are extracted according to the following process. In order to further distinguish the noise background from the gust front wind shear, first, the area with a combined shear value ≤2.5 m / (s·km) is removed. Since the radar detection signal has the characteristic of distance attenuation, the radial velocity data near the radar station has a relatively high signal-to-noise ratio, and strong precipitation and ground clutter may cause the combined shear value to be high, while far away from the radar station, the shear characteristics of the gust front are relatively weak due to signal attenuation, and are more easily covered by the noise background. Therefore, the judgment and extraction of the radial velocity convergence line / shear line related to the gust front follow the following standards: the gust front wind shear is divided into two categories. The first category is the strong wind shear category, which is usually close to the radar station (<30 km), and needs to meet the combined shear threshold ≥4 m / (s·km), the convergence line length ≥30 km, and the length-width ratio of the minimum bounding box ≥4:1; the second category is the weak wind shear category, which is distributed 30-100 km away from the station, the shear threshold is set to 2.5-4 m / (s·km), the convergence line length is 8-30 km, and the possible existence of fracture and other situations is allowed.

[0091] 2. Based on the gust front wind shear samples, coordinate system conversion, sample set division and data labeling are performed to obtain an expanded data set; specifically including:

[0092] 2.1 Convert the polar coordinate system of the gust front wind shear sample to the rectangular coordinate system, and then map the gust front wind shear sample to a 200x200 km, 588x588 resolution image to construct a strong wind shear data set and a weak wind shear data set.

[0093] 2.2 Extract 80% of the data in the strong wind shear data set and the weak wind shear data set respectively to divide the training set and the validation set, and then use the remaining 20% of the data as the prediction set.

[0094] 2.3 Use the LabelMe tool to dot-type label the gust front wind shear characteristics in the training set and the validation set to draw a closed polygon around the outer edge of the convergence line to obtain a labeled wind shear area, and based on the pixel coordinates of the closed curve, generate a JSON label file containing wind shear position information to obtain an expanded data set.

[0095] 3. Design a Mask R-CNN model architecture including a feature extraction module, a region candidate network module, a RoI Align module, and a multi-task output module, and train the Mask R-CNN model architecture to obtain the array wind front wind shear identification model;

[0096] 3.1 Mask R-CNN model architecture

[0097] As shown in Figure 3 , the overall model architecture design: the Mask R-CNN model adopts a modular architecture of "feature extraction-candidate region generation-precise alignment-multi-task output", which strengthens the learning of weak wind shear features on the basis of realizing target detection and instance segmentation.

[0098] 3.1.1 Backbone

[0099] As shown in Figure 4 , a Residual Network 50 (ResNet50) combined with a Feature Pyramid Networks (FPN) is used as a feature extraction module. The ResNet50 constructs a network by introducing a residual block. On the output layer of the convolution calculation of each module, an input item x is added directly to form a short circuit mechanism, and then the network calculation result F(x) is added to the input item x as the overall H(x) output. The short circuit mechanism helps to propagate the gradient and information transmission, thereby solving the gradient vanishing and overfitting problems in deep networks, ensuring that the subtle changes in the radial velocity data of the array wind front during movement can be accurately captured; the FPN combines different levels of feature pictures, including low-level high-resolution pictures and high-level rich semantic information pictures, to form new feature pictures, focusing on improving the recognition effect of weak wind shear.

[0100] 3.1.2 Region Proposals Network (RPN)

[0101] The main role of RPN is to generate different rectangular candidate boxes, that is, for each pixel point in the feature map, 9 candidate boxes of 3 different aspect ratios (1:1, 1:2, 2:1) and 3 different area sizes (128x128, 256x256, 512x512) are generated on the original image. After softmax classification (judging whether it is a squall shear) and boundary box regression (correcting the size and position of the candidate box), about 2000 initial candidate boxes are output. In order to reduce the number of candidate boxes, the Non-Maximum Suppression (NMS) algorithm is used, and the Intersection over Union (IoU) is set to 0.7, that is, when the candidate box overlaps with the correct area more than 0.7, it is given to retain, and finally 256 candidate regions (Proposals) are output.

[0102] 3.1.3Region of Interest Align (Roi Align)

[0103] To ensure that the candidate box is not offset when it is mapped to the feature map, and to avoid the deviation of the feature region corresponding to the candidate box from the region mapped to the feature map due to quantization. Roi Align uses bilinear interpolation to accurately map the generated candidate box to the feature image, and performs pooling on these feature images to finally obtain a distortion-free aligned feature map with a fixed size of 7x7, ensuring that the subtle feature information of the squall shear can be accurately preserved

[0104] 3.1.4 Multi-task output module

[0105] The multi-task output module includes a classification branch, a boundary box regression branch and a mask branch. The feature extraction module, the region candidate network module, the RoI Align module and the multi-task output module are combined to obtain the Mask R-CNN model architecture.

[0106] Classification branch: output the binary classification results of "squall shear" and "non-squall shear" through the fully connected layer (Fully Connected Layer), and use the cross-entropy loss function to optimize the classification accuracy.

[0107] Boundary box regression branch: output the coordinate parameters (t x, t y , t w , t h ) of the minimum bounding matrix of the squall shear, and minimize the deviation between the predicted box and the ground truth.

[0108] Mask branch: A 28x28 pixel binary mask is generated for each detected gust front wind shear by a Fully Convolution Nets layer, and a binary cross-entropy loss function is used to optimize the pixel-level segmentation accuracy.

[0109] 3.2 Model training

[0110] Model training scheme design: In order to make the model learn better strong and weak wind shear features and avoid overfitting during training, this study uses a phased training method and dynamically adjusts hyperparameters during training.

[0111] First phase: Mix the two types of training set samples in a 1:1 ratio, iterate with Batch Size=4 (each batch contains 2 strong wind shear samples and 2 weak wind shear samples), set the initial learning rate to 0.0025, use the stochastic gradient descent (SGD) optimizer, momentum parameter 0.9, and weight decay coefficient 0.0001. After 20 epochs of training, the validation set mAP_50 reaches more than 80%, and enters the second phase.

[0112] Second phase: To improve the recognition accuracy of weak wind shear samples, increase the proportion of weak wind shear samples to 1:3, and iterate with Batch Size=4 (each batch contains 1 strong wind shear sample and 3 weak wind shear samples), and the learning rate is decayed to 0.00025 to avoid overfitting. Through horizontal comparison of recognition results on the prediction set in different training periods (80, 100, and 120 epochs), the recognition accuracy is highest when iterating 100 epochs, and overfitting is likely to occur when iterating 120 epochs. Therefore, the model is trained for 100 epochs, and all weight parameters (.pth file) during training are saved, including the weight parameters of the backbone network, the weight parameters of the feature pyramid network, the weight parameters of the region candidate network, the weight parameters of Roi Align, the weight parameters of the multi-task output, and the weight parameters of the optimizer state.

[0113] (3) Model training effect evaluation: As shown in Figure 5 , the average precision mean (mAP_50) and total loss value are used to evaluate the model performance, where mAP_50 refers to the average precision obtained when the overlap between the predicted bounding box and the true bounding box is 0.5. At the same time, epoch represents the complete number of times the entire dataset is used to train the model. As the number of epochs increases, the learning effect of the model will continuously improve until the best performance of the model is reached, the training and optimization of the Mask R-CNN model architecture are completed, and the gust front wind shear recognition model is obtained.

[0114] As Figure 5 shown, the total loss value gradually approaches 0.1 with the increase of training cycles, which indicates that the model has strong learning ability in the process of continuous optimization and adaptation to training data. At the same time, the average precision mean (mAP_50) also increases rapidly and then tends to be stable, showing a small range of fluctuations, which reflects the gradual improvement of the model's performance in target recognition and positioning. After 20 training cycles, the mAP value of the squall front detection is basically stable above 0.8, which shows that the average recognition accuracy of the trained model on the validation set is above 80%.

[0115] 3.3 Evaluation indicators

[0116] The algorithm is evaluated using the confusion matrix. The confusion matrix combines the test set according to its true results and the model's predicted results into four categories: true positive (TP), the number of positive samples correctly predicted as positive by the model; false negative (FN), the number of positive samples incorrectly predicted as negative by the model; false positive (FP), the number of negative samples incorrectly predicted as positive by the model; and true negative (TN), the number of negative samples correctly predicted as negative by the model. Based on the confusion matrix, the following four key evaluation indicators are calculated: hit rate (POD), false alarm rate (FAR), missed alarm rate (MAR), and critical success index (CSI).

[0117] (10)

[0118] (11)

[0119] (12)

[0120] (13)

[0121] 3.4 Establishing a prediction data set: the remaining 20% of samples in the strong and weak wind shear data sets are used as an independent prediction set.

[0122] Prediction results: during the algorithm test process, the prediction set data is input, and the model predicts the data set by reading the weight parameter file. If there is a squall front wind shear, a black closed curve is drawn along the outer edge of the wind shear; if there is no wind shear, the original data set is output.

[0123] 4. Input the expanded data set into the squall front wind shear identification model for squall front wind shear detection.

[0124] 5. Test

[0125] In order to verify the universality of the algorithm, the present application uses an independent test set to test and evaluate the algorithm, and the evaluation results are shown in Table 1. The date, radar station name, and number of gust front wind shear samples of each gust front wind shear process are listed in Table 1, and the TP, FN and FP of each process are counted. The results show that the POD of the algorithm for the test set is 84.95%, the FAR is 2.47%, the MAR is 15.05%, and the CSI is 83.16%.

[0126] Table 1 Identification results of two types of gust front wind shear test sets

[0127] For the above method, the present application selects two gust front wind shear processes, corresponding to strong wind shear (case 1) and weak wind shear (case 5) respectively. From Figure 6 It can be seen from the above that the algorithm can better identify the characteristics of gust front wind shear.

[0128] Therefore, by providing a gust front wind shear identification method based on artificial intelligence, the present application can reduce the dependence on reflectivity factor data, and focus on constructing an identification model based on gust front radial velocity data, which can not only accurately identify gust front wind shear, but also realize pixel-level segmentation and positioning of the wind shear area, and improve the identification efficiency.

[0129] The embodiments in the specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts of each embodiment can be referred to each other.

[0130] The principles and implementation modes of the present application are described by applying specific examples in this paper. The above description of the embodiments is only used to help understand the method of the present application and its core idea; at the same time, for those skilled in the art, according to the idea of the present application, the specific implementation mode and application range will be changed. In conclusion, the content of the specification should not be understood as a limitation of the present application.

Claims

1. A gust front wind shear identification method based on artificial intelligence, characterized in that: include: Radial velocity data is collected using a new generation Doppler weather radar, and noise filtering, missing value supplementation, data smoothing, wind shear value calculation, and sample screening and extraction are performed on the radial velocity data to obtain gust front wind shear samples; Based on the gust front wind shear samples, coordinate system conversion, sample set division and data annotation are performed to obtain an expanded data set; Design a Mask R-CNN model architecture that includes a feature extraction module, a region proposal network module, a RoI Align module, and a multi-task output module, and train the Mask R-CNN model architecture to obtain a gust front wind shear recognition model; The expanded data set is input into the gust front wind shear recognition model to perform gust front wind shear detection.

2. The method for identifying gust front wind shear based on artificial intelligence according to claim 1, characterized in that: Radial velocity data is collected using a new generation Doppler weather radar. The radial velocity data is subjected to noise filtering, missing value supplementation, data smoothing, wind shear value calculation, and sample screening and extraction to obtain gust front wind shear samples, including: Using a new generation Doppler weather radar, the PPI scanning in VCP 21 mode was used, and radial velocity data at an elevation angle of 0.5° at an altitude of 1-2 km above the ground were selected. Based on the radial velocity data, a sliding window frequency method is used to filter out noise and supplement missing values, and a control variable method is used to test the key variable parameters of the sliding window to select the optimal parameter combination that is suitable for the radial velocity data; The radial velocity data is smoothed using a median filter method with a filter window of 3×3. When all data points in the filter window are valid values, the average value of the eight data points in the filter window is selected as the center point value until the full image filtering process is completed. Based on the radial velocity data, linearly fitting the velocity shear band using the least squares method to calculate radial shear and azimuthal shear, and combining the radial shear and the azimuthal shear to obtain a combined shear; According to the characteristics of radar physical parameters and the combined shear threshold, strong wind shear samples and weak wind shear samples in the radial velocity data are screened, and the characteristic areas of the strong wind shear samples and the weak wind shear samples are extracted to obtain gust front wind shear samples including the two types of samples.

3. The method for identifying gust front wind shear based on artificial intelligence according to claim 2, characterized in that: The calculation expression of the radial shear is: ； in, is the radial velocity, is the distance to the radar center, is the coefficient; The calculation expression of the azimuth shear is: ； in, is the azimuth; The calculation expression of the combined shear is: 。 4. The method for identifying gust front wind shear based on artificial intelligence according to claim 3, characterized in that: Based on the gust front wind shear samples, coordinate system conversion, sample set division and data annotation are performed to obtain an expanded data set, including: The polar coordinate system of the gust front wind shear sample is converted into a rectangular coordinate system, and then the gust front wind shear sample is mapped into an image with a resolution of 200×200 km and 588×588 to construct a strong wind shear dataset and a weak wind shear dataset; 80% of the data are extracted from the strong wind shear dataset and the weak wind shear dataset respectively to divide them into training sets and validation sets, and the remaining 20% ​​of the data are used as prediction sets; The LabelMe tool was used to perform dot annotation on the gust front wind shear features in the training set and the validation set, so as to draw a closed polygon around the outer edge of the convergence line to obtain the annotated wind shear area. Based on the pixel coordinates of the closed curve, a JSON label file containing the wind shear location information was generated to obtain the expanded dataset.

5. The method for identifying gust front wind shear based on artificial intelligence according to claim 4, characterized in that: Design the Mask R-CNN model architecture, which includes a feature extraction module, a region proposal network module, a RoI Align module, and a multi-task output module, including: Residual blocks and short-circuit mechanisms are introduced into ResNet50 to obtain a residual structure. Through the residual structure, directly connected input items are introduced into each convolutional calculation output layer of ResNet50 to output multi-level deep feature maps. Then, using the feature pyramid network, the low-level high-resolution features and high-level semantic features in the multi-level deep feature maps are fused to obtain multi-scale feature maps, completing the design of the feature extraction module. For each pixel in the multi-scale feature map, nine candidate boxes randomly combined with three different aspect ratios and three different scales are generated. Softmax classification is used to determine whether each candidate box contains the gust front wind shear target. The position and size of the candidate box are corrected to obtain the initial candidate box. The non-maximum suppression algorithm with an IoU threshold of 0.7 is then used to filter redundant boxes from the initial candidate box to obtain the candidate region, completing the design of the region candidate network module. The candidate region is mapped to the multi-scale feature map using bilinear interpolation, and the multi-scale feature map is pooled to obtain a 7×7 distortion-free alignment feature map, completing the design of the RoI Align module; A multi-task output module including a classification branch, a bounding box regression branch, and a mask branch is designed, and then the feature extraction module, the region candidate network module, the RoI Align module, and the multi-task output module are combined to obtain the Mask R-CNN model architecture.

6. The method for identifying gust front wind shear based on artificial intelligence according to claim 5, characterized in that: The classification branch is used to output a binary classification result of gust front wind shear and non-gust front wind shear through a fully connected layer, and optimize the binary classification result using a cross entropy loss function; The bounding box regression branch is used to output the coordinate parameters of the minimum circumscribed matrix of the gust front wind shear and minimize the deviation between the predicted box and the true box; The mask branch is used to generate a 28×28 pixel binary mask for each detected gust front wind shear through a fully convolutional network layer, and optimize the pixel-level segmentation accuracy using a binary cross entropy loss function.

7. The method for identifying gust front wind shear based on artificial intelligence according to claim 6, characterized in that: The Mask R-CNN model architecture is trained to obtain a gust front wind shear recognition model, including: The strong wind shear samples and weak wind shear samples in the training set are mixed in a ratio of 1:1, and 20 rounds of iterative training are performed with 4 samples as a batch. When the mAP_50 index of the validation set is greater than or equal to 80%, the first stage of training is completed and the second stage of training begins; each batch includes 2 strong wind shear samples and 2 weak wind shear samples; After the first phase of training, the ratio of the weak wind shear samples was increased to 1:3, the learning rate was reduced, and different training cycles were compared horizontally. The weight parameters during all training periods were saved to obtain the model parameters and complete the second phase of training. After the second phase of training, the mean average precision and total loss value are evaluated on the validation set based on the model parameters to obtain the training evaluation results, completing the training optimization of the Mask R-CNN model architecture and obtaining the gust front wind shear recognition model; The confusion matrix is ​​used to calculate the evaluation index to evaluate the detection results of the gust front wind shear recognition model and complete the model optimization.

8. The method for identifying gust front wind shear based on artificial intelligence according to claim 7, characterized in that: The confusion matrix includes true positives, false positives, false negatives and true negatives, and the evaluation indicators include hit rate, false detection rate, missed detection rate and critical success index.

Images