A method and system for aircraft wake vortex identification

Abstract

This invention discloses a method and system for identifying aircraft wake vortices, belonging to the field of aerospace technology. The method involves acquiring radial velocity field data of aircraft wake vortices scanned by a lidar RHI (Radio Hierarchical Induction) radar; converting the radial velocity field data into a color image; using a ResNet18 network pre-trained on the ImageNet large-scale natural image dataset as a trained visual feature encoder, with all model weights frozen; extracting feature vectors from the color image based on the trained visual feature encoder; training a classifier using the feature vectors as input and binary classification labels (vortex presence or absence) as output; constructing a recognition network model based on the trained visual feature encoder and classifier; converting the radial velocity field data of the aircraft wake vortex to be tested into a corresponding color image, inputting it into the recognition network model, and outputting the recognition result (vortex presence or absence). This method can fundamentally alleviate the overfitting risk inherent in end-to-end training under small sample conditions.

Description

Technical Field

[0001] This invention relates to the field of aviation technology, and more specifically to a method and system for identifying aircraft wake vortices. Background Technology

[0002] During flight, the pressure difference between the upper and lower surfaces of an aircraft wing creates a pair of strong, counter-rotating vortices at the wingtip, known as the aircraft wake vortex. These wake vortices can persist for several minutes within the airport approach airspace. If a subsequent aircraft enters the wake vortex region of the preceding aircraft, it will experience a strong induced roll moment, seriously endangering flight safety. To address this, the International Civil Aviation Organization (ICAO) has established static wake separation standards based on aircraft weight classification, requiring a certain time or distance interval between aircraft. However, current standards are conservative and fail to dynamically reflect the actual dissipation state of the wake vortex, thus limiting airport operational efficiency. Therefore, achieving real-time and accurate wake vortex identification has significant engineering application value for improving airport capacity and ensuring flight safety.

[0003] Currently, end-to-end deep learning methods based on convolutional neural networks (CNNs) have become the mainstream technology in the field of vortex identification. This method renders the radial velocity field data from a range-height indicator (RHI) acquired by LiDAR into a two-dimensional image, directly inputting it into a pre-trained convolutional neural network (such as VGG, ResNet, YOLO, etc.) for end-to-end training. Through multi-layer convolution and pooling operations, it automatically extracts deep spatial features from the velocity field image, thereby achieving binary classification of vortex presence / absence or precise localization of vortex location. This method does not require manual feature design, theoretically possesses powerful feature representation capabilities, and can achieve high recognition accuracy when sufficient training data is available.

[0004] However, in actual airport lidar monitoring scenarios, the generation of aircraft wake vortices is random and transient, resulting in a small proportion of vortex samples in the total daily scan data. Furthermore, the high cost of manual annotation leads to a severe shortage of available training samples. Under these conditions, directly training a deep convolutional network with a massive number of parameters from scratch, far exceeding the available sample size, makes the model highly susceptible to overfitting. This manifests as excellent performance on the training set but a significant drop in performance on the test set, indicating poor generalization ability. Summary of the Invention

[0005] To address the problems existing in the above-mentioned fields, this invention proposes an aircraft wake vortex identification method and system. The constructed identification network model explicitly decouples deep feature extraction from classifier training, enabling the model to inherit the powerful feature expression capabilities of convolutional networks while minimizing the scale of parameters that need to be learned from labeled data. This fundamentally alleviates the inherent overfitting risk of end-to-end training under small sample conditions.

[0006] To address the aforementioned technical problems, this invention discloses a method for identifying aircraft wake vortices, comprising the following steps: Acquire radial velocity field data of the aircraft wake vortex from LiDAR RHI scans; convert the radial velocity field data into a color image; A ResNet18 network pre-trained on the ImageNet large-scale natural image dataset will be used as the trained visual feature encoder, and all model weights will be frozen. The feature vector of the color image is extracted based on the trained visual feature encoder. The feature vector is used as input and the binary classification label with or without vortices is used as output to train the classifier and obtain the trained classifier. Based on the trained visual feature encoder and classifier, a recognition network model is constructed; After converting the radial velocity field data of the aircraft wake vortex into a color image, the image is input into the recognition network model, and the output is the recognition result of whether there is a vortex or not.

[0007] Preferably, the feature vector of the color image is extracted based on the trained visual feature encoder, specifically including: The visual feature encoder extracts feature vectors from color images through a global average pooling layer; Specifically, the global average pooling layer is the pool5 layer in the ResNet18 network, located before the fully connected classification head.

[0008] Preferably, the classifier is a radial basis function kernel support vector machine.

[0009] Preferably, the step of training a classifier by taking feature vectors as input and binary classification labels (with or without vortices) as output to obtain the trained classifier specifically includes: Normalize the feature vectors; A stratified sampling strategy is used to divide the normalized feature vector into training sets according to a set ratio. The ratio of vortex samples to non-vortex samples in the training set is consistent with the original LiDAR dataset of the aircraft tail vortex. The radial basis function kernel support vector machine is trained using the training set to obtain the optimal weights of the radial basis function kernel support vector machine and determine the trained classifier.

[0010] Preferably, the step of constructing a recognition network model based on the trained visual feature encoder and classifier specifically includes: The trained classifier, visual feature encoder, and the size parameters of the input color image are saved together to construct the recognition network model.

[0011] Preferably, the step of converting the radial velocity field data of the aircraft wake vortex to be measured into a color image, inputting it into the recognition network model, and outputting the recognition result of whether there is a vortex or not specifically includes: When the identification result is vortex, extract and output the corresponding vortex cycle number; When the classifier outputs a recognition result of no vortex, the recognition result of the current scanning cycle is marked as no vortex, and the subsequent processing steps for that scanning cycle are skipped, and the process directly proceeds to the next scanning cycle.

[0012] Preferably, acquiring the radial velocity field data of the aircraft wake vortex obtained by lidar RHI scanning specifically includes: Obtain the raw lidar dataset of the aircraft wake vortex, which includes radial velocity and beam elevation angle; The original dataset is grouped in one complete RHI scan cycle, invalid values ​​in the radial velocity are replaced with NaN, and the polar coordinates are transformed to rectangular coordinates according to the elevation angle of each beam to generate a two-dimensional radial velocity field matrix.

[0013] Preferably, the step of converting the radial velocity field data into a color image specifically includes: The radial velocity field data were resampled to a uniform grid using bilinear interpolation. Using jet color scales, within a preset speed mapping range, a uniform grid is rendered as a standardized color image of preset pixels through off-screen rendering.

[0014] Preferably, the color image is a color image that meets the input size requirements of the ResNet18 network.

[0015] Preferably, it also includes an aircraft wake vortex identification system, comprising: The data acquisition module is used to acquire radial velocity field data of the aircraft wake vortex from the LiDAR RHI scan; and to convert the radial velocity field data into a color image. The model building module is used to take the ResNet18 network pre-trained on the ImageNet large-scale natural image dataset as the trained visual feature encoder and freeze all model weights; based on the trained visual feature encoder, feature vectors of color images are extracted, and the classifier is trained with the feature vectors as input and binary classification labels of vortex or no vortex as output to obtain the trained classifier; based on the trained visual feature encoder and classifier, a recognition network model is constructed. The wake vortex identification module is used to convert the radial velocity field data of the aircraft wake vortex under test into a color image, input it into the identification network model, and output the identification result of whether there is a vortex or not.

[0016] Compared with the prior art, the present invention has the following beneficial effects: The proposed aircraft wake vortex identification method utilizes a network model that freezes pre-trained weights. The model employs the general visual feature extraction capabilities learned from training a ResNet18 network on the ImageNet dataset as a fixed visual feature encoder, directly applying it to the wake vortex velocity field. This achieves efficient cross-domain knowledge transfer. Simultaneously, the visual feature encoder automatically and deeply mines high-order semantic information from the velocity field image, accurately capturing the spatial topological relationship between positive and negative vortex velocities and the local gradient morphology at the vortex core edge. This automatically extracted structural features possess more fundamental discriminative power than manually extracted statistical features. The feature vectors extracted by the visual feature encoder are used as input to the classifier, with binary labels indicating whether a vortex is present or absent as output. This training of the classifier allows for more focused and robust classification of boundaries, effectively mitigating class imbalance issues in end-to-end deep learning. This identification network model architecture explicitly decouples the visual feature encoder from the classifier training, essentially reducing the complex deep network to a pure feature computation graph, eliminating the need for backpropagation. The actual training is handled solely by the lightweight classifier. This decoupling not only accelerates the training process, but also eliminates the need to rely on a large deep learning training framework when performing batch inference on edge devices. This significantly improves the stability of the engineering process, allowing the model to inherit the powerful feature representation capabilities of convolutional networks while minimizing the size of parameters that need to be learned from labeled data. This fundamentally alleviates the inherent risk of overfitting in end-to-end training under small sample conditions. Attached Figure Description

[0017] Figure 1 This is a flowchart of the aircraft wake vortex identification method proposed in this invention; Figure 2 This is a schematic diagram of a Doppler lidar scan provided in an embodiment of the present invention; Figure 3 A schematic diagram illustrating the specific implementation steps of the aircraft wake vortex identification method provided in this embodiment of the invention; Figure 4 A color image converted from radial velocity field data provided in an embodiment of the present invention. Detailed Implementation

[0018] The following will refer to the appendices in the embodiments of the present invention. Figures 1-4 The technical solutions in the embodiments of the present invention will be clearly and completely described. It should be understood that the terminology used in the present invention is only for describing particular implementation methods and is not intended to limit the present invention.

[0019] Example like Figure 1As shown, this invention proposes a method for identifying aircraft wake vortices, which includes the following steps: S1: Acquire radial velocity field data of the aircraft wake vortex from the LiDAR RHI scan; convert the radial velocity field data into a color image; S2: A ResNet18 network pre-trained on the ImageNet large-scale natural image dataset will be used as the trained visual feature encoder, and all model weights will be frozen. S3: Extract feature vectors from color images based on the trained visual feature encoder. Use the feature vectors as input and binary classification labels with or without vortices as output to train the classifier and obtain the trained classifier. S4: Construct a recognition network model based on the trained visual feature encoder and classifier; S5: After converting the radial velocity field data of the aircraft wake vortex to be measured into a color image, input it into the recognition network model and output the recognition result of whether there is a vortex or not.

[0020] Specifically, in step S1, the raw lidar dataset of the aircraft's wake vortex is acquired. This raw lidar dataset includes information such as the radial velocity matrix (RadVrData, with 52 range gates) and beam elevation angle (BeamEl). Figure 2 The image shown is a schematic diagram of a Doppler lidar scan. In this embodiment, the elevation angle of the laser radar's transmitted beam relative to the horizontal plane is defined as follows: The beams increase sequentially within the range of 0° to 20°, and each RHI scan cycle contains a total of 52 beams with different elevation angles; O 1 and O 2 are the center positions of the left and right vortex cores of the aircraft tail vortex, respectively. The left and right vortices rotate in opposite directions, together forming a typical counter-rotating vortex pair structure. r c The radius of the vortex core represents the radial scale of the solid rotation region of the vortex core, and is used in the Hallock-Burnham vortex model to describe the velocity distribution characteristics inside the vortex core. b 0 represents the aircraft's wingspan. r The slant range is the distance along the line of sight of the lidar beam, i.e., the radial distance from the laser pulse along the beam propagation direction to the target detection point; v r is the radial velocity component measured by the lidar along the line of sight of the beam, and is the projection of the tangential velocity of the wake vortex onto the line of sight of the lidar. The direction towards the lidar is positive, and the direction away from the lidar is negative. v t The tangential velocity of the wake vortex at the detection point is given by the Hallock-Burnham vortex model based on the vortex core circulation Γ and the vortex core radius.r c The distance from the detection point to the vortex core is determined together; delta The angle between the direction of the tangential velocity of the wake vortex at the detection point and the line-of-sight direction of the lidar beam determines the tangential velocity. v t The projection ratio in the line of sight, i.e. v r = v t ·cos( delta In this embodiment, the original dataset is grouped into groups of 52 consecutive beam records as a complete RHI scan cycle. Invalid values ​​(±999m / s) in the radial velocity are replaced with NaN, and the polar coordinates are transformed to rectangular coordinates according to the elevation angle of each beam to generate a two-dimensional radial velocity field matrix with horizontal distance as the horizontal axis and height as the vertical axis for subsequent interpolation and rendering.

[0021] Because the original radar scan data is not uniformly distributed in a Cartesian coordinate system, such as Figure 3 As shown, the method proposed in this embodiment first uses the input RADV format lidar binary file as the raw data; then, it performs data parsing and coordinate transformation on the raw data, including beam grouping, invalid value processing, and conversion from polar coordinates to rectangular coordinates.

[0022] Subsequently, the griddata bilinear interpolation method was used to perform velocity field difference and image rendering, and the two-dimensional radial velocity field matrix of each period was resampled to a 100×100 uniform grid. Using the jet color scale, within the preset velocity mapping range (±5m / s), the uniform grid was rendered into a standardized color image of 224×224 pixels through off-screen rendering.

[0023] like Figure 4 The image shown is a color image converted from radial velocity field data. The red areas correspond to positive radial velocities towards the lidar, while the blue areas correspond to negative radial velocities away from the lidar. A distinct red-and-blue alternating pair of positive and negative velocities is visible in the center of the image, concentrated in a region with a horizontal distance of 400–600 m and a height of 30–120 m, consistent with the typical spatial distribution characteristics of an aircraft wake vortex during approach. This spatial topology of the positive and negative velocity pairs is the core criterion for identifying deep convolutional features extracted from the network model in this invention.

[0024] The rendering process is performed off-screen, meaning that image generation is completed without relying on a graphical user interface, effectively avoiding interaction interference with the application interface during batch processing. The velocity mapping range is selected based on the typical velocity magnitude of the wake vortex during the approach stage in the measured data. This range ensures sufficient color contrast between positive and negative velocities for the structure in the image, while suppressing visual interference from the background wind field.

[0025] In step S2, a ResNet18 network pre-trained on the ImageNet large-scale natural image dataset, with all its weights frozen, is used as the visual feature encoder for the constructed recognition network model. The visual feature encoder is followed by a classifier, which employs a radial basis function kernel support vector machine.

[0026] In step S3, a standardized 224×224 pixel color image is input into a visual feature encoder, and a 512-dimensional feature vector of the color image is extracted through a global average pooling layer. The global average pooling layer is the pool5 layer in the ResNet18 network, located before the fully connected classification head.

[0027] The selection of the pool5 layer is based on the following: This layer precedes the fully connected classification head, and its output feature vector is an abstraction that has undergone multiple convolutional and pooling operations. It effectively represents the high-level semantic information of the velocity field image (such as the spatial distribution of positive and negative velocity pairs, the local gradient structure at the edge of the vortex kernel, etc.). Furthermore, its compact 512-dimensional representation has sufficient discriminative information density, making it suitable as input to subsequent classifiers. During inference, all weights of the ResNet18 backbone remain fixed (frozen) and do not participate in any parameter updates. This ensures that the feature extraction process relies entirely on the general image representation capabilities accumulated through ImageNet pre-training, rather than relearning from limited wake vortex labeled samples.

[0028] Before training, the 512-dimensional feature vectors are normalized to eliminate the dimensional differences between different feature dimensions.

[0029] A stratified sampling strategy was adopted to divide the normalized feature vectors into a training set (including 807 training samples, of which about 209 samples have positive vortexes) and a test set at a set ratio of 7:3. This ensured that the ratio of samples with vortices to samples without vortices in the training set and the test set was consistent with the original LiDAR dataset of the aircraft tail vortex, in order to address the inherent class imbalance problem between samples with vortex periods (about 25%) and samples without vortex periods (about 75%) in the dataset.

[0030] The constructed recognition network model includes a classifier connected after the visual feature encoder. This classifier employs a radial basis function (RBF) kernel support vector machine (SVM). The RBF kernel SVM is trained using the training set to obtain its optimal weights and determine the trained classifier.

[0031] In step S4, after training is completed, the trained classifier, visual feature encoder, and the size parameters of the input color image are saved in a unified manner to construct the recognition network model, which can be directly loaded and called in the inference stage without repeated training.

[0032] In step S5, the radial velocity field data of the aircraft wake vortex to be measured is converted into a color image and then input into the recognition network model cycle by cycle according to the scanning cycle. When the classifier outputs a recognition result indicating a vortex, the corresponding vortex cycle number is extracted and output. When the classifier outputs a recognition result of "no vortex", the recognition result of the current scanning cycle is marked as "no vortex" and the subsequent processing steps for that scanning cycle are skipped, and the process directly proceeds to the next RHI scanning cycle.

[0033] This embodiment uses four metrics—Accuracy (ACC), Precision Per Count (PPV), True Rate of Return (TPR), and F1 score—to comprehensively evaluate the performance of the constructed recognition network model. The definitions of each metric are as follows: ACC = (TP + TN) / (TP + TN + FP + FN); PPV = TP / (TP + FP); TPR = TP / (TP + FN); F1 = 2 × PPV × TPR / (PPV + TPR); TP, TN, FP, and FN represent the number of true positives, true negatives, false positives, and false negatives, respectively.

[0034] Under the above dataset partitioning conditions, the performance metrics of the training set and test set of the method proposed in this invention are shown in Table 1, and the confusion matrix of the test set is shown in Table 2.

[0035] Table 1 Performance metrics for training and testing sets Table 2 Test Set Confusion Matrix Analysis of the data in Table 1 shows that the recognition network model constructed in this invention achieves a recognition accuracy of 97.0% on the test set and a vortex recall (TPR) of 88.2%. Analysis of the data in Table 2 shows that only 2 cases were missed (FN=2). It is worth noting that the performance on the test set is slightly better than that on the training set, indicating that the model does not exhibit overfitting and possesses good generalization ability, thus verifying the effectiveness of the transfer learning strategy under small sample conditions.

[0036] This embodiment demonstrates that the recognition network model constructed by the method proposed in this invention achieves a test set accuracy of 97.0%, a vortex recall rate of 88.2%, and an overall F1 score of 81.1% under the condition of only 807 training samples (of which approximately 209 are positive vortex samples). It is worth noting that the test set performance is slightly better than the training set (95.2%), indicating that the model has good generalization ability and does not exhibit overfitting. This result verifies the effectiveness of the transfer learning strategy—by freezing the backbone weights of the ResNet18 network and training only the lightweight SVM classification head, the constructed recognition network model inherits the general image representation capabilities of the pre-trained network while minimizing the scale of parameters that need to be learned from labeled data, thus achieving stable and reliable recognition performance under the realistic constraint of scarce labeled samples of aviation wake vortices.

[0037] For a wake vortex safety early warning system, the cost of a missed alarm (misjudging a vortex period as a non-vortex period) is far greater than a false alarm (misjudging a non-vortex period as a vortex period). The missed alarm result of FN=2 indicates that this method has high reliability in practical engineering applications.

[0038] Compared with the manual feature support vector machine method, the recognition network model constructed in this invention replaces manual statistical features with deep convolutional features, which can automatically capture the spatial topology of positive and negative velocity pairs in the velocity field and the local gradient information near the vortex kernel. It fundamentally breaks through the performance limit of manual features in scenarios with strong background wind field interference, and is suitable for direct transfer and application in different detection scenarios.

[0039] At the engineering deployment level, this invention fully integrates the identification method with the functions of raw LiDAR data parsing, velocity field visualization, cycle-by-cycle result browsing, and result export. It constructs a fully automated processing link from raw RADV format data input to identification result output. Users do not need to have programming skills to complete all operations from data loading to risk assessment, which significantly reduces the threshold for using the method.

[0040] This invention also proposes an aircraft wake vortex identification system, comprising: The data acquisition module is used to acquire radial velocity field data of the aircraft wake vortex from the LiDAR RHI scan; and to convert the radial velocity field data into a color image. The model building module is used to take the ResNet18 network pre-trained on the ImageNet large-scale natural image dataset as the trained visual feature encoder and freeze all model weights; based on the trained visual feature encoder, feature vectors of color images are extracted, and the classifier is trained with the feature vectors as input and binary classification labels of vortex or no vortex as output to obtain the trained classifier; based on the trained visual feature encoder and classifier, a recognition network model is constructed. The wake vortex identification module is used to convert the radial velocity field data of the aircraft wake vortex under test into a color image, input it into the identification network model, and output the identification result of whether there is a vortex or not.

[0041] The core objective of the proposed method is to construct a recognition method that achieves an effective balance between feature representation capability and small-sample generalization. This invention proposes rendering the radial velocity field data of a LiDAR RHI (Radio High Intensity Index) radar into a color image. The constructed recognition network model utilizes a ResNet18 network, fully pre-trained on the large-scale ImageNet dataset, for deep feature extraction. Furthermore, by freezing the weights of the backbone network, a lightweight SVM classifier head is trained using only the 512-dimensional feature vectors of the extracted color image. This model architecture explicitly decouples deep feature extraction from classifier training, allowing the model to inherit the powerful feature representation capabilities of convolutional networks while minimizing the number of parameters that need to be learned from labeled data. This fundamentally alleviates the overfitting risk inherent in end-to-end training under small-sample conditions.

[0042] The method proposed in this invention fully integrates the functions of parsing the acquired raw binary data of lidar, visualizing the velocity field, interactively browsing the results cycle by cycle, and batch exporting. It constructs a fully automated processing link from raw RADV format data to the output of recognition results, enabling users without a programming background to deploy and use it directly.

[0043] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

[0044] Furthermore, unless otherwise stated, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. All references to this specification are incorporated by way of citation to disclose and describe methods relating to those references. In the event of any conflict with any incorporated reference, the content of this specification shall prevail.

Claims

1. A method for identifying aircraft wake vortices, characterized in that, Includes the following steps: Acquire radial velocity field data of the aircraft wake vortex from LiDAR RHI scans; convert the radial velocity field data into a color image; A ResNet18 network pre-trained on the ImageNet large-scale natural image dataset will be used as the trained visual feature encoder, and all model weights will be frozen. The feature vector of the color image is extracted based on the trained visual feature encoder. The feature vector is used as input and the binary classification label with or without vortices is used as output to train the classifier and obtain the trained classifier. Based on the trained visual feature encoder and classifier, a recognition network model is constructed; After converting the radial velocity field data of the aircraft wake vortex into a color image, the image is input into the recognition network model, and the output is the recognition result of whether there is a vortex or not.

2. The aircraft wake vortex identification method according to claim 1, characterized in that, The extraction of feature vectors from color images based on the trained visual feature encoder specifically includes: The visual feature encoder extracts feature vectors from color images through a global average pooling layer; Specifically, the global average pooling layer is the pool5 layer in the ResNet18 network, located before the fully connected classification head.

3. The aircraft wake vortex identification method according to claim 1, characterized in that, The classifier uses a radial basis function kernel support vector machine.

4. The aircraft wake vortex identification method according to claim 1, characterized in that, The process of training a classifier by taking feature vectors as input and binary labels (with or without vortices) as output, to obtain the trained classifier, specifically includes: Normalize the feature vectors; A stratified sampling strategy is used to divide the normalized feature vector into training sets according to a set ratio. The ratio of vortex samples to non-vortex samples in the training set is consistent with the original LiDAR dataset of the aircraft tail vortex. The radial basis function kernel support vector machine is trained using the training set to obtain the optimal weights of the radial basis function kernel support vector machine and determine the trained classifier.

5. The aircraft wake vortex identification method according to claim 1, characterized in that, The step of constructing a recognition network model based on the trained visual feature encoder and classifier specifically includes: The trained classifier, visual feature encoder, and the size parameters of the input color image are saved together to construct the recognition network model.

6. The aircraft wake vortex identification method according to claim 1, characterized in that, The process of converting the radial velocity field data of the aircraft wake vortex into a color image, inputting it into the recognition network model, and outputting the recognition result of whether or not a vortex exists includes: When the identification result is vortex, extract and output the corresponding vortex cycle number; When the classifier outputs a recognition result of no vortex, the recognition result of the current scanning cycle is marked as no vortex, and the subsequent processing steps for that scanning cycle are skipped, and the process directly proceeds to the next scanning cycle.

7. The aircraft wake vortex identification method according to claim 1, characterized in that, The acquisition of radial velocity field data of the aircraft wake vortex obtained from lidar RHI scanning specifically includes: Obtain the raw lidar dataset of the aircraft wake vortex, which includes radial velocity and beam elevation angle; The original dataset is grouped in one complete RHI scan cycle, invalid values ​​in the radial velocity are replaced with NaN, and the polar coordinates are transformed to rectangular coordinates according to the elevation angle of each beam to generate a two-dimensional radial velocity field matrix.

8. The aircraft wake vortex identification method according to claim 1, characterized in that, The process of converting radial velocity field data into a color image specifically includes: The radial velocity field data were resampled to a uniform grid using bilinear interpolation. Using jet color scales, within a preset speed mapping range, a uniform grid is rendered as a standardized color image of preset pixels through off-screen rendering.

9. The aircraft wake vortex identification method according to claim 1, characterized in that, The color image is a color image that meets the input size requirements of the ResNet18 network.

10. An aircraft wake vortex identification system, characterized in that, include: The data acquisition module is used to acquire radial velocity field data of the aircraft wake vortex from the LiDAR RHI scan; and to convert the radial velocity field data into a color image. The model building module is used to take the ResNet18 network pre-trained on the ImageNet large-scale natural image dataset as the trained visual feature encoder and freeze all model weights; based on the trained visual feature encoder, feature vectors of color images are extracted, and the classifier is trained with the feature vectors as input and binary classification labels of vortex or no vortex as output to obtain the trained classifier; based on the trained visual feature encoder and classifier, a recognition network model is constructed. The wake vortex identification module is used to convert the radial velocity field data of the aircraft wake vortex under test into a color image, input it into the identification network model, and output the identification result of whether there is a vortex or not.

Images