An artificial intelligence-based gust front wind shear identification method

Abstract

This invention provides an artificial intelligence-based method for identifying gust front wind shear, comprising: performing noise filtering, missing value imputation, data smoothing, wind shear value calculation, and sample selection and extraction on collected radial velocity data to obtain gust front wind shear samples; based on the gust front wind shear samples, performing coordinate system transformation, sample set partitioning, and data annotation to obtain an expanded dataset; designing and training a Mask R-CNN model architecture to obtain a gust front wind shear identification model; and inputting the expanded dataset into the gust front wind shear identification model for gust front wind shear detection. This invention reduces the dependence on reflectivity factor data, focusing on building the identification model based on gust front radial velocity data, which can not only accurately identify gust front wind shear but also achieve pixel-level segmentation and localization of wind shear regions, improving identification efficiency.

Description

Technical Field

[0001] This invention relates to the field of gust front shear recognition technology, and in particular to an artificial intelligence-based gust front shear recognition method. Background Technology

[0002] Gust fronts are small- to medium-scale hazardous weather phenomena, currently primarily monitored using Doppler weather radar. Gust front wind shear manifests as banded wind convergence or wind shear regions in the radar velocity field. Existing gust front wind shear identification algorithms mainly fall into two categories: one based on narrowband echo features in radar reflectivity factor images, such as bidirectional gradient methods and dual-template local binarization segmentation; the other based on the radial velocity features of the gust front, such as the detection algorithm proposed by Uyeda, which achieves wind shear identification by fusing radial convergence and azimuth shear features, combined with threshold filtering and feature classification. However, this method suffers from a high false alarm rate and insufficient robustness in complex environments. Later, Lincoln Laboratory developed a gust front wind shear algorithm based on fuzzy logic algorithms and image processing techniques, which significantly improved identification accuracy and enabled operational applications, but it requires high-quality radar data and has certain limitations in its applicability.

[0003] Most domestic and international solutions underutilize radial velocity, or can only achieve coarse-grained qualitative detection of wind shear regions, resulting in inaccurate segmentation and localization, cumbersome processes, and difficulties in business automation. In recent years, deep learning technology has been increasingly applied to the intelligent identification of severe weather. However, using deep learning to identify gust front wind shear has certain limitations: firstly, most methods primarily use the reflectivity factor data of the gust front, with low utilization of radial velocity data, which is closely related to gust front wind shear; secondly, some algorithms can only achieve coarse localization of wind shear regions, failing to achieve pixel-level precise segmentation, and the processes are complex, leading to numerous problems in practical business applications. Summary of the Invention

[0004] To overcome the shortcomings of existing technologies, the purpose of this invention is to provide an artificial intelligence-based method for identifying gust front wind shear. This method reduces the reliance on reflectivity factor data and focuses on building an identification model based on gust front radial velocity data. It can not only accurately identify gust front wind shear but also achieve pixel-level segmentation and positioning of the wind shear region, thereby improving identification efficiency.

[0005] To achieve the above objectives, the present invention provides the following solution: a method for identifying gust front shear based on artificial intelligence, comprising:

[0006] Radial velocity data was collected using a new generation of Doppler weather radar, and noise was filtered out, missing values ​​were supplemented, data was smoothed, wind shear values ​​were calculated, and samples were screened and extracted to obtain gust front wind shear samples.

[0007] Based on the aforementioned gust front shear samples, coordinate system transformation, sample set partitioning, and data labeling are performed to obtain an expanded dataset.

[0008] 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 a gust front shear recognition model.

[0009] The expanded dataset is input into the gust front shear recognition model to perform gust front shear detection.

[0010] Optionally, radial velocity data is acquired using a new generation of Doppler weather radar, and the radial velocity data is subjected to noise filtering, missing value compensation, data smoothing, wind shear value calculation, and sample selection and extraction to obtain gust front wind shear samples, including:

[0011] Using a new generation of Doppler weather radar, PPI scanning was performed in VCP 21 mode, and radial velocity data at an elevation angle of 0.5° at an altitude of 1-2 km above the ground were selected.

[0012] Based on the radial velocity data, noise filtering and missing value supplementation were performed using the sliding window frequency method, and the key variable parameters of the sliding window were tested using the controlled variable method to screen the optimal parameter combination that fits the radial velocity data.

[0013] The radial velocity data is smoothed using a median filtering method with a 3×3 filtering window. 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 filtering of the entire image is completed.

[0014] Based on the radial velocity data, the velocity shear band is linearly fitted using the least squares method to calculate the radial shear and azimuth shear. The radial shear and the azimuth shear are then combined to obtain the combined shear.

[0015] Based on the radar physical parameter characteristics and combined shear threshold, strong wind shear samples and 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 gust front shear samples including the two types of samples.

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

[0017] ;

[0018] in, Radial velocity, Distance to the radar center For coefficients;

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

[0020] ;

[0021] in, It is the azimuth angle;

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

[0023] .

[0024] Optionally, based on the gust front shear samples, coordinate system transformation, sample set partitioning, and data labeling are performed to obtain an expanded dataset, including:

[0025] The polar coordinate system of the gust front shear sample is converted into a rectangular coordinate system, and then the gust front shear sample is mapped into an image with a resolution of 200×200km and 588×588 to construct a strong wind shear dataset and a weak wind shear dataset.

[0026] 80% of the data is extracted from the strong wind shear dataset and the weak wind shear dataset respectively to divide the dataset into training set and validation set, and the remaining 20% ​​of the data is used as 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 closed polygons around the outer edge of the convergence line, thereby obtaining the marked wind shear regions. Based on the pixel coordinates of the closed curves, a JSON tag file containing wind shear location information is generated to obtain the expanded dataset.

[0028] Optionally, the Mask R-CNN model architecture can be designed, including a feature extraction module, a region candidate network module, a RoI Align module, and a multi-task output module, comprising:

[0029] Residual blocks and short-circuit mechanisms are introduced into ResNet50 to obtain a residual structure. Through the residual structure, directly connected input terms are introduced into the output layer of each convolutional computation of ResNet50 to output multi-level deep feature maps. Then, the feature pyramid network is used to fuse the low-level high-resolution features and high-level semantic features in the multi-level deep feature maps to obtain multi-scale feature maps, thus completing the design of the feature extraction module.

[0030] For each pixel in the multi-scale feature map, nine candidate boxes are generated by randomly combining three different aspect ratios and three different scales. Softmax classification is used to determine whether each candidate box contains gust front shear targets to correct the position and size of the candidate boxes and obtain initial candidate boxes. Then, a non-maximum suppression algorithm with an IoU threshold of 0.7 is used to filter redundant boxes in the initial candidate boxes to obtain candidate regions, thus completing the design of the region candidate network module.

[0031] The candidate region is mapped onto the multi-scale feature map using bilinear interpolation, and the multi-scale feature map is pooled to obtain a 7×7 distortion-free aligned feature map, thus completing the design of the RoI Align module.

[0032] The design includes a multi-task output module comprising a classification branch, a bounding 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 then combined to obtain the Mask R-CNN model architecture.

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

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

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

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

[0037] The strong wind shear samples and the weak wind shear samples in the training set are mixed in a 1:1 ratio, and 20 rounds of iterative training are performed with 4 samples per 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.

[0038] After the first phase of 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 cycles are compared horizontally. The weight parameters during all training periods are saved to obtain the model parameters and complete the second phase of training.

[0039] After the second phase of training is completed, based on the model parameters, the mean accuracy and total loss are evaluated on the validation set to obtain the training evaluation results. The training optimization of the Mask R-CNN model architecture is completed, and the gust wind shear recognition model is obtained.

[0040] Evaluation metrics are calculated using a confusion matrix to assess the effectiveness of the gust front shear identification model and optimize the model.

[0041] Optionally, the confusion matrix includes true positives, false positives, false negatives, and true negatives, and the evaluation metrics include hit rate, false positive rate, false negative rate, and critical success index.

[0042] This invention discloses the following technical effects by providing an artificial intelligence-based method for identifying gust front shear:

[0043] 1) In data acquisition and preprocessing: It can effectively eliminate complex noise in radar data and use sliding window statistics to achieve adaptive filtering and supplementary measurements. It can also incorporate richer multi-source meteorological data fusion, such as multi-level elevation angles and more radar physical quantities. It significantly reduces the dependence on radar reflectivity factor data and focuses on using radial velocity data closely related to gust front shear, thereby improving the accuracy and reliability of gust front shear identification.

[0044] 2) In terms of data-label set construction: the unification of polar-Cartesian coordinate transformation and high resolution provides an excellent data foundation for the adaptation of backend deep learning algorithms. Classified training is beneficial for balancing the model's generalization ability.

[0045] 3) Model training: By implementing an advanced integrated framework for object detection and segmentation, the gradient flow of deep networks can be optimized. By dynamically adjusting hyperparameters in multiple stages, the ratio of strong to weak wind shear samples can be dynamically switched, improving the model's ability to learn small sample-weak features, thereby improving the overall recognition accuracy.

[0046] 4) Model evaluation and result output: By using closed curve segmentation and pixel-level mask prediction, the system ensures accurate positioning, comprehensive evaluation dimensions, stable prediction, and high overall recognition accuracy. This allows the invention to expand the test set to cover more extreme and complex weather conditions.

[0047] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0048] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

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

[0050] Figure 2 A flowchart illustrating the method architecture provided in this embodiment of the invention;

[0051] Figure 3 A flowchart of an automatic gust front shear recognition algorithm based on Mask R-CNN provided in an embodiment of the present invention;

[0052] Figure 4 This is a schematic diagram of the residual block structure provided in an embodiment of the present invention;

[0053] Figure 5 The graphs showing the variation of average accuracy and total loss with training epochs are provided for embodiments of the present invention.

[0054] Figure 6 The radial velocity identification results at five typical moments during two gust front shear processes provided in this embodiment of the invention are shown in the figure. Detailed Implementation

[0055] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0056] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0057] like Figure 1 As shown, this invention provides an artificial intelligence-based method for identifying gust front shear, comprising:

[0058] 1. Radial velocity data is acquired using a new generation of Doppler weather radar, and the radial velocity data undergoes noise filtering, missing value supplementation, data smoothing, wind shear calculation, and sample selection and extraction to obtain gust front wind shear samples; specifically including:

[0059] 1.1 Using a new generation of Doppler weather radar, PPI scanning was performed in VCP 21 mode. This mode acquires data such as echo intensity and radial velocity within the radar's detection range through horizontal circular scanning. Since gust front shear typically occurs at an altitude of 1-2 km above the ground, radial velocity data from the 0.5° elevation angle layer were selected for preprocessing.

[0060] 1.2 Noise filtering and missing measurement value compensation

[0061] Based on the radial velocity data, noise filtering and missing value supplementation were performed using the sliding window frequency method, and the key variable parameters of the sliding window were tested using the controlled variable method to screen the optimal parameter combination that fits the radial velocity data.

[0062] The principle of the "sliding window frequency method" is as follows: A sliding window containing an M×N distance database is set up. The radial velocity data at a 0.5° elevation angle layer in PPI mode is traversed through this window. The center of the window is considered the distance database to be processed, and the processing result is determined by the statistical analysis of all distance database values ​​within the window. The specific method is as follows:

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

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

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

[0066] Zc= (1)

[0067] In the formula, k1 is the noise filtering threshold, k2 is the threshold for supplementing missing measurements, and "ND" indicates no valid data. Formula (1) utilizes the principle of spatial continuity 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 It is valid data, and when F n+1When k ≤ k1, it indicates 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 satisfies n+1 When the value is greater than k1, it indicates that there is a large amount of data near the center of the window, which should be retained. For missing values, when the value of the center point of the window is Z... ij When F is ND, if max If k ≥ 2, it means that a large number of data points in the same interval were detected around the center of the window. Therefore, the median value Z of the interval with the highest frequency can be used. m Assign it a complement value; conversely, if F max If the value is less than k2, it is assumed that there is relatively little valid data around the center of the window, and therefore no further adjustments are needed for Z. ij Perform the value replacement.

[0068] The values ​​of M, N, n, λΔd, k1, and k2 in formula (1) have a certain impact on the preprocessing results. Therefore, it is necessary to conduct controlled variable experiments to compare the effects of noise filtering and missing measurement supplementation under different parameter conditions, and select the optimal parameter configuration for data preprocessing. The optimal parameter configuration obtained by this paper after experimentation is: M×N=3×3, n=15, λ=2, Δd=4, k1=3, k2=4.

[0069] 1.3 Data Smoothing

[0070] To eliminate data pulsation, a median filtering method with a 3×3 filtering window 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 until the whole image filtering process is completed.

[0071] 1.4 Calculation of Wind Shear Value Based on Least Squares Method

[0072] To characterize the geometry of the radial velocity convergence / shear line of the gust front, a linear fit of the velocity shear band is performed using the least squares method based on the radial velocity data. Radial shear is defined as the variation of radar radial velocity along the radial direction with distance. A "fitting window" is selected to calculate the radial shear, which must include n data points on the same radial direction, i.e., (v1, r1), (v... i , r i )...(v n , r n (v) i Represents radial velocity, r i (This represents the distance to the radar center). Let the regression equation be:

[0073] (2)

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

[0075] (3)

[0076] To obtain the best-fit line, the sum of squared relative errors calculated above must be minimized, that is, the value of D in the following formula must be minimized:

[0077] (4)

[0078] have to:

[0079] (5)

[0080] (6)

[0081] Solving for:

[0082] (7)

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

[0084] Azimuth shear is defined as the radial velocity changing clockwise from 0° to 360° along a range circle at the same distance from the radar. The calculation process is the same as for radial shear, and the azimuth shear can be obtained as follows:

[0085] (8)

[0086] The combined shear is obtained by combining the radial shear and the azimuth shear:

[0087] (9)

[0088] 1.5 Determination and Extraction of Wind Shear Lines

[0089] Based on radar physical parameter characteristics and combined shear thresholds, strong wind shear samples and weak wind shear samples are filtered from the radial velocity data, and feature regions of the strong wind shear samples and weak wind shear samples are extracted to obtain gust front shear samples including both types of samples. Specifically, this includes:

[0090] To more accurately identify gust front wind shear, this invention filters gust front samples based on radar physical parameter characteristics. These characteristics include: a "narrow-band echo" feature in the reflectivity factor map; convergence lines or shear lines in the radial velocity map; significantly high spectral width values ​​in the spectral width map; regular movement characteristics; and the presence of parent thunderstorms that trigger the gust front in its vicinity. Based on the above characteristics, the acquired radar data is filtered, and samples meeting the criteria are extracted according to the following process. To further distinguish between noise background and gust front wind shear, areas with combined shear values ​​≤2.5 m / (s·km) are first eliminated. Due to the distance attenuation characteristic of radar detection signals, radial velocity data near the radar station has a higher signal-to-noise ratio. Heavy precipitation and ground clutter may cause the combined shear value to be artificially inflated. However, when the gust front is far from the radar station, the shear characteristics are relatively weak due to signal attenuation and are more easily masked by noise background. Therefore, the determination and extraction of radial velocity convergence lines / shear lines related to gust fronts follow the following criteria: Gust front wind shear is divided into two categories. The first category is strong wind shear, which is usually close to the radar station (<30 km), and must meet the following requirements: combined shear threshold ≥4 m / (s·km), convergence line length ≥30 km, and the aspect ratio of the minimum circumscribed matrix ≥4:1; the second category is weak wind shear, which is distributed 30-100 km away from the station, with a shear threshold set at 2.5-4 m / (s·km), convergence line length 8-30 km, and the possibility of breakage is allowed.

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

[0092] 2.1 The polar coordinate system of the gust front shear sample is converted into a rectangular coordinate system, and then the gust front shear sample is mapped into an image with a resolution of 200×200km and 588×588 to construct a strong wind shear dataset and a weak wind shear dataset.

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

[0094] 2.3 Using the LabelMe tool, the gust front wind shear features in the training set and the validation set are marked with dots to draw closed polygons around the outer edge of the convergence line, thereby obtaining the marked wind shear regions. Based on the pixel coordinates of the closed curves, a JSON tag file containing wind shear location information is generated to obtain the expanded dataset.

[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 a gust front shear recognition model;

[0096] 3.1 Mask R-CNN Model Architecture

[0097] like Figure 3 As shown, the overall architecture design of the model is as follows: The Mask R-CNN model adopts a modular architecture of "feature extraction - candidate region generation - precise alignment - multi-task output". On the basis of achieving object detection and instance segmentation, it strengthens the learning of weak wind shear features.

[0098] 3.1.1 Backbone

[0099] like Figure 4 As shown, Residual Network50 (ResNet50) combined with Feature Pyramid Networks (FPN) is used as the feature extraction module. ResNet50 constructs the network by introducing residual blocks. At the output layer of each module's convolutional computation, a directly connected input term x is added, forming a short-circuit mechanism. The network computation result F(x) is then added to the input term x to produce the overall output H(x). This short-circuit mechanism facilitates gradient propagation and information transfer, thus solving the gradient vanishing and overfitting problems in deep networks and ensuring that subtle changes in the radial velocity data of gust fronts during their movement are accurately captured. FPN, on the other hand, combines feature images from different levels, including low-level high-resolution images and high-level images rich in semantic information, to form new feature images, focusing on improving the recognition effect of weak wind shear.

[0100] 3.1.2 Region Proposals Network (RPN)

[0101] The main function of RPN is to generate different rectangular candidate boxes. Specifically, it generates nine candidate boxes for each pixel in the feature map, with three different aspect ratios (1:1, 1:2, 2:1) and three different area sizes (128×128, 256×256, 512×512). After softmax classification (to determine if it represents a gust of wind or frontal shear) and bounding box regression (to correct the size and position of the candidate boxes), approximately 2000 initial candidate boxes are output. To reduce the number of candidate boxes, a non-maximum suppression (NMS) algorithm is used, setting the intersection over union (IoU) ratio to 0.7. This means that candidate boxes with an overlap greater than 0.7 with the correct region are retained, ultimately outputting 256 candidate regions (Proposals).

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

[0103] To ensure that candidate regions are not offset when mapped to feature maps, Roi Align avoids deviations between the feature regions corresponding to candidate boxes and the regions mapped to the feature maps due to quantization. Roi Align uses bilinear interpolation to accurately map the generated candidate boxes onto the feature images and then pools these feature images to obtain a fixed-size 7×7 distortion-free aligned feature map, ensuring that subtle feature information of gust front shear is accurately preserved.

[0104] 3.1.4 Multi-task output module

[0105] The Mask R-CNN model architecture is obtained by combining the multi-task output module, which includes a classification branch, a bounding box regression branch, and a mask branch, and then combining the feature extraction module, the region candidate network module, the RoI Align module, and the multi-task output module.

[0106] Classification branch: The fully connected layer outputs binary classification results of "is it a gust front shear" and "is not a gust front shear", and the cross-entropy loss function is used to optimize the classification accuracy.

[0107] Bounding box regression branch: Outputs the coordinate parameters (t) of the minimum bounding matrix of the gust front shear. x, t y , t w , t h It employs a method that minimizes the deviation between the predicted bounding box and the ground truth.

[0108] Masking branch: A 28×28 pixel binary mask is generated for each detected gust front shear using fully convolutional network layers, and the pixel-level segmentation accuracy is optimized using a binary cross-entropy loss function.

[0109] 3.2 Model Training

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

[0111] Phase 1: The two training sets are mixed in a 1:1 ratio and iterated with a batch size of 4 (each batch contains 2 strong wind shear samples and 2 weak wind shear samples). The initial learning rate is set to 0.0025, and a stochastic gradient descent (SGD) optimizer is used with a momentum parameter of 0.9 and a weight decay coefficient of 0.0001. After training for 20 epochs, the validation set mAP_50 reaches over 80%, and the process proceeds to Phase 2.

[0112] The second stage: To improve the recognition accuracy of weak wind shear samples, the ratio of weak wind shear samples was increased to 1:3. Iterations were also performed with a batch size of 4 (each batch containing 1 strong wind shear sample and 3 weak wind shear samples), and the learning rate was decayed to 0.00025 to avoid overfitting. By comparing the recognition results on the prediction set with different training periods (80, 100, and 120 epochs), it was found that the recognition accuracy was highest with 100 epochs, while overfitting was more likely with 120 epochs. Therefore, this invention selected 100 epochs as the model's training period and saved all weight parameters (.pth file) during training, including the weight parameters of the backbone network, the feature pyramid network, the region candidate network, Roi Align, the multi-task output, and the optimizer state.

[0113] (3) Evaluation of model training performance: such as Figure 5 As shown, this study uses the mean average accuracy (mAP_50) and total loss to evaluate model performance. mAP_50 refers to the average accuracy obtained when the overlap between the predicted bounding box and the ground truth bounding box is 0.5. Meanwhile, epoch represents the total number of times the entire dataset has been used to train the model. With the increase of the number of epochs, the model's learning effect typically improves until it reaches its optimal performance, thus completing the training and optimization of the Mask R-CNN model architecture and obtaining the gust wind shear recognition model.

[0114] like Figure 5 As shown, the total loss value gradually approaches 0.1 with the increase of training cycles, indicating that the model has a strong learning ability in the process of continuously optimizing and adapting to the training data. Meanwhile, the mean accuracy (mAP_50) also increases rapidly before stabilizing, showing fluctuations within a small range. This phenomenon reflects the gradual improvement of the model's performance in target recognition and localization. After 20 training cycles, the mAP value for gust front detection is basically stable above 0.8, indicating that the trained model achieves an average recognition accuracy of over 80% on the validation set.

[0115] 3.3 Evaluation Indicators

[0116] The algorithm was evaluated using a confusion matrix. The confusion matrix grouped the test set into four categories based on the true results and the model's predictions: True Positive (TP), the number of positive samples correctly predicted as positive; False Negative (FN), the number of positive samples incorrectly predicted as negative; False Positive (FP), the number of negative samples incorrectly predicted as positive; and True Negative (TN), the number of negative samples correctly predicted as negative. Based on the confusion matrix, four key performance indicators (KPIs) were calculated: Point of Detection (POD), False Algorithm Rate (FAR), False Negative Rate (MAR), and Critical Success Index (CSI).

[0117] (10)

[0118] (11)

[0119] (12)

[0120] (13)

[0121] 3.4 Establish the prediction dataset: Use the remaining 20% ​​of the samples in the strong and weak wind shear datasets as an independent prediction set.

[0122] Prediction Results: During algorithm testing, the input prediction set data is used, and the model predicts the dataset by reading the weight parameter file. If a gust wind shear exists, a black closed curve is drawn along the outer edge of the wind shear; if no wind shear exists, the original dataset is output.

[0123] 4. Input the expanded dataset into the gust front shear recognition model to perform gust front shear detection.

[0124] 5. Testing

[0125] To verify the universality of the algorithm, this invention uses an independent test set to test and evaluate the algorithm, and the evaluation results are shown in Table 1. Table 1 lists the date of occurrence, radar station name, and number of gust front shear samples for each gust front shear process, and also calculates the TP, FN, and FP for each process. The results show that the algorithm achieves a POD of 84.95%, FAR of 2.47%, MAR of 15.05%, and CSI of 83.16% on the test set.

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

[0127]

[0128] To illustrate the above method, this invention selected two gust front wind shear processes, corresponding to strong wind shear (Example 1) and weak wind shear (Example 5) respectively, for demonstration. From Figure 6 As can be seen, the algorithm can identify the characteristics of gust fronts and wind shears quite well.

[0129] Therefore, this invention provides an artificial intelligence-based method for identifying gust front wind shear, which reduces the reliance on reflectivity factor data and focuses on building an identification model based on gust front radial velocity data. This method can not only accurately identify gust front wind shear, but also achieve pixel-level segmentation and positioning of the wind shear region, thereby improving identification efficiency.

[0130] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0131] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. An artificial intelligence-based gust front wind shear identification method, characterized in that, include: Radial velocity data was acquired using a new generation of Doppler weather radar. Noise filtering, missing value imputation, data smoothing, wind shear calculation, and sample selection were then performed on the radial velocity data to obtain gust front wind shear samples, including: Using a new generation of Doppler weather radar, PPI scanning was performed in VCP 21 mode, 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, noise filtering and missing value supplementation were performed using the sliding window frequency method, and the key variable parameters of the sliding window were tested using the controlled variable method to screen the optimal parameter combination that fits the radial velocity data. The radial velocity data is smoothed using a median filtering method with a 3×3 filtering window. 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 filtering of the entire image is completed. Based on the radial velocity data, the velocity shear band is linearly fitted using the least squares method to calculate the radial shear and azimuth shear. The radial shear and the azimuth shear are then combined to obtain the combined shear. The calculation expression for the radial shear is: ； in, Radial velocity, Distance to the radar center For coefficients; The calculation expression for the azimuth shear is: ； wherein is an azimuth angle; The calculation expression for the combined shear is: ； Based on the radar physical parameter characteristics and combined shear threshold, strong wind shear samples and 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 gust front shear samples including the two types of samples. Based on the aforementioned gust front shear samples, coordinate system transformation, sample set partitioning, and data labeling are performed to obtain an expanded dataset. 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 a gust front shear recognition model. The expanded dataset is input into the gust front shear recognition model to perform gust front shear detection.

2. The method of claim 1, wherein the method is based on artificial intelligence. Based on the aforementioned gust front shear samples, coordinate system transformation, sample set partitioning, and data labeling are performed to obtain an expanded dataset, including: The polar coordinate system of the gust front shear sample is converted into a rectangular coordinate system, and then the gust front shear sample is mapped into an image with a resolution of 200×200km and 588×588 to construct a strong wind shear dataset and a weak wind shear dataset. 80% of the data is extracted from the strong wind shear dataset and the weak wind shear dataset respectively to divide the dataset into training set and validation set, and the remaining 20% ​​of the data is used as prediction set. Using the LabelMe tool, the gust front wind shear features in the training set and the validation set are marked with dots to draw closed polygons around the outer edge of the convergence line, thereby obtaining the marked wind shear regions. Based on the pixel coordinates of the closed curves, a JSON tag file containing wind shear location information is generated to obtain the expanded dataset. 3.The method of claim 2, wherein, The Mask R-CNN model architecture is designed, including a feature extraction module, a region candidate network module, a RoI Align module, and a multi-task output module, comprising: Residual blocks and short-circuit mechanisms are introduced into ResNet50 to obtain a residual structure. Through the residual structure, directly connected input terms are introduced into the output layer of each convolutional computation of ResNet50 to output multi-level deep feature maps. Then, the feature pyramid network is used to fuse the low-level high-resolution features and high-level semantic features in the multi-level deep feature maps to obtain multi-scale feature maps, thus completing the design of the feature extraction module. For each pixel in the multi-scale feature map, nine candidate boxes are generated by randomly combining three different aspect ratios and three different scales. Softmax classification is used to determine whether each candidate box contains gust front shear targets to correct the position and size of the candidate boxes and obtain initial candidate boxes. Then, a non-maximum suppression algorithm with an IoU threshold of 0.7 is used to filter redundant boxes in the initial candidate boxes to obtain candidate regions, thus completing the design of the region candidate network module. The candidate region is mapped onto the multi-scale feature map using bilinear interpolation, and the multi-scale feature map is pooled to obtain a 7×7 distortion-free aligned feature map, thus completing the design of the RoI Align module. The design includes a multi-task output module comprising a classification branch, a bounding 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 then combined to obtain the Mask R-CNN model architecture.

4. The method for identifying gust front shear based on artificial intelligence according to claim 3, characterized in that: The classification branch is used to pass through a fully connected layer to output a binary classification result of gust front shear and non-gust front shear, and to optimize the binary classification result using the cross-entropy loss function. The bounding box regression branch is used to output the coordinate parameters of the minimum outer matrix of the gust front 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 shear through a fully convolutional network layer, and to optimize pixel-level segmentation accuracy using a binary cross-entropy loss function.

5. The method of claim 4, wherein the method is based on artificial intelligence. The Mask R-CNN model architecture is trained to obtain a gust front shear recognition model, including: The strong wind shear samples and the weak wind shear samples in the training set are mixed in a 1:1 ratio, and 20 rounds of iterative training are performed with 4 samples per 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 is completed, the proportion of the weak wind shear class samples is increased to 1:3, the learning rate is reduced, and different training cycles are compared horizontally. The weight parameters during all training periods are saved to obtain the model parameters and complete the second phase of training. After the second phase of training is completed, based on the model parameters, the mean accuracy and total loss are evaluated on the validation set to obtain the training evaluation results. The training optimization of the Mask R-CNN model architecture is completed, and the gust wind shear recognition model is obtained. Evaluation metrics are calculated using a confusion matrix to assess the effectiveness of the gust front shear identification model and optimize the model.

6. The method of claim 5, wherein the method is based on artificial intelligence. The confusion matrix includes true positives, false positives, false negatives, and true negatives, and the evaluation metrics include hit rate, false positive rate, false negative rate, and critical success index.

Images