A method and system for real-time identification and warning of aircraft wake vortex through multi-source data fusion

Abstract

The application relates to the technical field of civil aviation operation safety and intelligent air traffic control, and discloses a multi-source data fusion aircraft wake vortex real-time identification and alarm method and system. Multi-source data is collected through various radars, and then unified timestamp alignment is performed on the source data to obtain a wake vortex data set; a parallel space-time deep learning wake vortex identification model is constructed, time domain features are extracted through a gate cycle unit, and spatial domain features are extracted through a convolutional neural network; the time domain features and the spatial domain features are fused, and finally the identification result and the existence probability of the wake vortex are predicted; dynamic probability threshold is combined for judgment to suppress false positives and false negatives; finally, the wake vortex identification result and the judgment result are generated, and are pushed to the terminal for output, prompting and log recording. The application realizes real-time and reliable identification of the wake vortex under complex weather and multi-aircraft conditions, improves the identification accuracy and robustness, and reduces false positives and false negatives.

Description

Technical Field

[0001] This invention relates to the field of civil aviation operation safety and intelligent air traffic control technology, specifically to a method and system for real-time identification and warning of aircraft wake vortices through multi-source data fusion. Background Technology

[0002] Wake vortex hazards originate from wingtip vortices, creating rolling moments and aerodynamic disturbances on following aircraft, necessitating strict separation management in the terminal area. Airport operations are typically organized based on visual and meteorological estimates or fixed separation rules according to aircraft type (Heavy / Medium / Light).

[0003] Among them, Doppler coherent lidar can measure radial wind vectors through coherent detection principle, and RHI mode can detect wingtip vortices by performing profile scanning along the direction perpendicular to the flight path; wind profiler radar provides background wind field and atmospheric stability information; ground-based LiDAR / radar and surveillance systems can provide three-dimensional flight paths of aircraft (latitude / longitude, distance, altitude, heading / ground speed, etc.).

[0004] Current methods for wake vortex detection and identification based on single radar data struggle to integrate with multi-source information such as flight paths and meteorological data, failing to form a complete and accurate spatiotemporal background wind field and wake vortex field. Relying solely on aircraft type and fixed interval operating rules cannot reflect real-time wake vortex intensity and dissipation rate, leading to safety redundancy or insufficient efficiency. Furthermore, identification networks based on single-channel time series or single-frame images struggle to simultaneously capture temporal dependencies and spatial structural features.

[0005] The temporal and spatial coordinates of multi-source data (ground-based radar, dual-coherent Doppler lidar A / B, wind profiler radar) are difficult to align uniformly, making direct data stitching and modeling challenging. Wake vortex observations exhibit both strong temporal and spatial structural characteristics, which traditional methods struggle to adequately model, resulting in models that fail to adequately represent the joint temporal and spatial structures, thus limiting robustness and generalization ability.

[0006] The probability uncertainty of the identification results is high under different weather and aircraft operating conditions. The threshold setting lacks adaptability and dynamic thresholds based on operating conditions (stability, background wind, aircraft type, and operating stage). It is difficult to balance false alarms and false negatives. Summary of the Invention

[0007] To address the aforementioned problems, the present invention aims to provide a method and system for real-time identification and warning of aircraft wake vortices based on multi-source data fusion. It constructs an engineering-applicable multi-source data fusion and online identification scheme, achieving real-time and reliable wake vortex identification under complex weather conditions and with multiple aircraft types, while improving identification accuracy and robustness, and reducing false alarms and missed alarms. The technical solution is as follows:

[0008] A method for real-time identification and alarm of aircraft wake vortices through multi-source data fusion, comprising the following steps:

[0009] Step 1: Multi-source data acquisition: Multi-source data is acquired through airport ground-based LiDAR / radar, runway edge wind sensors, two Doppler coherent lidars (A and B), and wind profiler radar.

[0010] Step 2: Spatiotemporal alignment and fusion: Perform unified timestamp alignment on the data from the airport ground-based LiDAR / radar, Doppler laser coherent radar, and wind profiler radar to obtain the wake vortex dataset;

[0011] Step 3: Wake Vortex Identification: Construct a parallel spatiotemporal deep learning wake vortex identification model, including gated recurrent units and convolutional neural networks; preprocess the wake vortex dataset, dividing it into two input parts: sequence input. and block input Input the sequence Input-gated recurrent units extract time-domain features and block inputs. The input is fed into a convolutional neural network to extract spatial domain features; the temporal domain features and spatial domain features are then fused to predict the identification result and existence probability of the wake vortex.

[0012] Step 4: Probability and Dynamic Threshold Strategy: Combine dynamic probability thresholds for judgment to suppress false alarms and false negatives;

[0013] Step 5: Alarm generation, push and recording: Generate wake vortex identification results and judgment results, push the output to the tower and unit to each terminal for prompts and record the log.

[0014] A real-time aircraft wake vortex identification and warning system based on multi-source data fusion includes:

[0015] Acquisition module: Multi-source data acquisition is performed through airport ground-based LiDAR / radar, runway edge wind sensor, two Doppler coherent lidars (A and B), and wind profiler radar;

[0016] Alignment module: Performs unified timestamp alignment on the data from various sources, including airport ground-based LiDAR / radar, Doppler laser coherent radar, and wind profiler radar, to obtain the wake vortex dataset;

[0017] Recognition Module: Constructs a parallel spatiotemporal deep learning wake vortex recognition model, including gated recurrent units and convolutional neural networks; preprocesses the wake vortex dataset, dividing it into two input parts: sequence input. and block input Input the sequence Input-gated recurrent units extract time-domain features and block inputs. The input is fed into a convolutional neural network to extract spatial domain features; the temporal domain features and spatial domain features are then fused to predict the identification result and existence probability of the wake vortex.

[0018] Alarm module: Combines dynamic probability thresholds for judgment to suppress false alarms and missed alarms;

[0019] Output module: alarm generation, push and logging: generates wake vortex identification results and judgment results, pushes them to the control tower and the unit, and logs them to each terminal.

[0020] The beneficial effects of this invention are:

[0021] 1) This invention achieves multi-source fusion and spatiotemporal alignment through unified multi-source data interface access, anomaly removal and consistency checks, as well as spatiotemporal alignment and data block construction in the runway coordinate system, which significantly reduces false detections / missed detections caused by single-source deviations and coordinate mismatches.

[0022] 2) This invention employs a parallel spatiotemporal network, with bidirectional GRU extracting temporal features and CNN extracting spatial features. Gated weighted fusion and classification enhance the joint modeling capability of wake vortex temporal evolution and spatial structure, thereby improving robustness and generalization under complex conditions.

[0023] 3) This invention adopts a dynamic threshold strategy that combines stability / aircraft type / stage to adapt to different stability, wind field and aircraft type operating conditions, and achieves a balance between low false alarms and low false alarms based on dual radar consistency, duration event generation mechanism and hierarchical alarm. Attached Figure Description

[0024] Figure 1 This is a flowchart of a method for real-time identification and alarm of aircraft wake vortices based on multi-source data fusion, according to an embodiment of the present invention.

[0025] Figure 2 This is a schematic diagram of an A / B lidar scan of a wind field.

[0026] Figure 3 This is a diagram illustrating the overall framework for wake vortex identification in an embodiment of the present invention.

[0027] Figure 4 This image illustrates the results of using a parallel spatiotemporal deep learning network to identify wake vortices in wind field data acquired by lidar.

[0028] Figure 5 This is a system architecture diagram according to an embodiment of the present invention.

[0029] Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0030] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0031] This invention relates to the field of civil aviation operation safety and intelligent air traffic control technology, specifically to a method and system for real-time identification and warning of aircraft wake vortices based on multi-source data fusion. The method flowchart is shown below. Figure 1 As shown, the method includes the following steps: First, the spatial position information of the aircraft, the wake vortex data generated by the aircraft, and airport meteorological information are acquired separately through airport ground-based radar, two nearby Doppler coherent lidars, and wind profiler radar (including environmental information sensors such as temperature and humidity). Simultaneously, datasets are constructed through normalization preprocessing and spatiotemporal alignment. Then, a parallel spatiotemporal deep learning network is used to serialize and block the synthesized dual-channel data from the A / B construction. The sequence input and block input are respectively extracted and fused using GRU (Gated Recurrent Unit) and CNN (Convolutional Neural Network), and then trained through a deep learning neural network to predict the wake vortex identification result and its probability of existence. Subsequently, a dynamic threshold is used for judgment to suppress false alarms and missed alarms. Finally, the identification result and judgment result are generated and pushed to the control tower and aircraft crew, and output to each terminal for prompting and logging. This method can achieve real-time and reliable wake vortex identification under complex weather and multi-aircraft conditions, reducing false alarms and missed alarms, and improving the safe operation of the airport terminal area.

[0032] Step S1: Multi-source data acquisition: First, multi-source data is acquired through airport ground-based radar, runway edge wind sensors, Doppler lidar, and wind profiler radar.

[0033] like Figure 2As shown, two adjacent Doppler coherent lidars (hereinafter referred to as A / B) with a vertical distance not exceeding 10m are based on the coherent detection principle. They use frequency-stabilized pulsed laser illumination to receive the scattered echoes from wind-drifting aerosols in the atmosphere and coherently mix them with the local oscillator light. The radial wind vector in the line-of-sight direction of the laser beam is obtained by digital frequency discrimination of the intermediate frequency signal. It is essential that the scanning cycles of Doppler coherent lidars A and B are spaced half a cycle apart to ensure that the scanning directions of the two lidars are opposite at any given time (even if one is scanning clockwise and the other counterclockwise). The selected Doppler coherent lidars A and B have advantages such as high precision, speed, and rich data products, and can detect the low-to-mid-altitude and surrounding wind fields of the airport in real time. During the detection process, two Doppler coherent lidars, A and B, used range height indication (RHI) scanning mode to perform profile scanning along the direction perpendicular to the aircraft's flight path, forming a real-time wind field dataset. The radar measurements included timestamps and positional parameters (range, azimuth, elevation), effectively monitoring wingtip vortices. Additionally, a wind profiler radar simultaneously acquired meteorological elements (wind speed, wind direction, temperature, humidity, boundary layer stability, and turbulence intensity, etc.) near the airport and radar, used to obtain background wind field and atmospheric stability data. Connecting to the airport's ground-based radar and surveillance system allowed for the acquisition of three-dimensional flight path information (latitude / longitude, distance, altitude, heading / ground speed, etc.) for approaching and departing aircraft. All data was integrated through a unified interface and underwent anomaly removal and consistency checks, providing reliable input for subsequent spatiotemporal alignment and fusion.

[0034] Step S2: Spatiotemporal alignment and fusion. Perform unified timestamp alignment on all data sources, including airport ground-based LiDAR (Light Detection and Ranging Doppler Laser Detector) / radar, Doppler lidar, and wind profiler radar.

[0035] Aircraft tracks acquired by airport-based LiDAR / radar are transformed from a geographic coordinate system to a runway coordinate system referenced to the runway centerline. Resampling and interpolation are then performed on a unified time step and spatial grid, resulting in spatiotemporally consistent data blocks. Furthermore, considering the influence of the background wind field, the wind field input data from the two Doppler coherent lidars, A and B, are normalized to facilitate training and recognition by the parallel deep learning network. The calculation expression is as follows:

[0036] (1);

[0037] In the formula, Wind field data acquired by radar, σ and σ' are the mean and standard deviation calculated within a given time window or spatial neighborhood, respectively. This is the standardized result, i.e., the wake vortex dataset used for wake vortex identification.

[0038] Step S3: Wake Vortex Identification. A parallel spatiotemporal deep learning wake vortex identification model is constructed, including gated recurrent units and convolutional neural networks; the wake vortex dataset is preprocessed, consisting of two input parts: sequence input. and block input Input the sequence Input-gated recurrent units extract time-domain features and block inputs. The input is fed into a convolutional neural network to extract spatial domain features; then the temporal and spatial domain features are fused to finally predict the identification result and existence probability of the wake vortex.

[0039] Since both Doppler coherent light lidars, A and B, use periodic multi-discrete point scanning, when the two lidars are placed on the same side, their detection areas are relatively close, allowing their respective wind field information to be used as independent input channels, with feature fusion automatically learned by a deep learning network.

[0040] The fused complete wind field data (dual-channel data) serves as the input X (i.e., "radar-acquired wind field data") for normalization in formula (1). Subsequently, as described in step S2, the fused wind field data X is normalized to obtain the standardized result Y. Wake vortex scanning detection inherently possesses temporal and spatial characteristics. Therefore, the Doppler coherent lidar output Y can be equivalently considered as a scanning sequence at different times. A matrix composed of different blocks This invention proposes a parallel spatiotemporal deep learning wake vortex identification model that employs branching processing. A GRU and a CNN respectively extract the temporal and spatial features of the wind field detected by A / B Doppler coherent lidar, and then fuse these features at the feature layer before inputting them into a classifier for identification. Specifically, the wake vortex dataset is first preprocessed, divided into two input parts: sequence input SI and block input PIN. These two parts are then input into the GRU and CNN modules, respectively. The GRU extracts temporal features, and the CNN extracts spatial features. After feature extraction, the two types of features are fused, and finally, a fully connected layer is used to obtain the prediction result. The overall framework is as follows: Figure 3 As shown, a parallel spatiotemporal deep learning network is employed. The wind field sequence input from A / B lidar scanning is used to extract temporal features via a GRU. Simultaneously, the wind field data is segmented into blocks, which are then input into a CNN to extract spatial features. Finally, feature layer fusion is performed to output the wake vortex identification result and its probability of existence.

[0041] GRU temporal feature extraction: In a coherent lidar system that scans the detection area cyclically from neighboring angles, the echo sequence data exhibits significant temporal dependence. GRU can be used to extract the temporal features of radar echo data at different scanning angles within a single scan area. The GRU unit is input at time t. Compared to the previous hidden state As input, the hidden state at time t is output by updating and resetting the gates to control information forgetting and introducing. While maintaining long-term dependency modeling capabilities, it requires fewer parameters and exhibits better training stability, making it suitable for online inference and deployment. A bidirectional GRU is employed to bidirectionally encode and learn representations of the time-domain information from two Doppler coherent lidars, A and B, thereby focusing on the complete past and future contextual information of each point in the input sequence. For example... Figure 3 As shown, in this embodiment, an N-layer bidirectional GRU network processes the radial velocity wind field, comprising forward and backward sequences with multiple individual hidden layers, which are then fed forward and fused into the same output layer. The hidden layer vectors of the forward and backward input characteristics are obtained through learning. and :

[0042] , (2);

[0043] In the formula, For GRU mapping operators; and These represent the standardization results. The data is used as forward and backward inputs for sequential data. Then, the hidden layer vectors from the forward and backward processes are concatenated (CONCAT) to obtain the fused features, which are calculated as follows:

[0044] (3);

[0045] In the formula, This is a time-domain fusion feature.

[0046] Spatial Domain Feature Extraction using CNN: Spatial domain features of the radial velocity flow field of the wake vortex obtained from two coherent light lidars (A and B) are extracted using a CNN. Previous research has shown that dividing the image matrix into multiple patches is key to extracting local features of spatial information. Therefore, the wind field data is first divided into multiple patches, and then a CNN is applied to this set of patches. The convolution process is represented as follows:

[0047] (4);

[0048] In the formula, To represent the feature map output by the convolutional layer, This represents the set of convolutional kernels with output channels c. This represents an element-wise non-linear activation function (such as ReLU or Sigmoid). Subsequent convolutional layers are followed by activation functions to generate spatial domain feature maps. Spatial domain feature output. for:

[0049] (5);

[0050] In the formula, This is a feature extraction function based on a convolutional neural network. The function includes several convolutional layers, activation layers, and optional normalization layers, used to extract features from the input feature map. Spatial domain features are extracted from them.

[0051] Spatiotemporal feature fusion: A gated weighted fusion module is used to first map the wake temporal and spatial features to a consistent dimension.

[0052] (6);

[0053] In the formula, and For learnable parameters, , They are respectively , Uniform dimensional features after linear mapping.

[0054] Then, the two feature streams are concatenated and input into the gating unit to obtain the gating vector. :

[0055] (7);

[0056] In the formula, and For learnable parameters, This represents the concatenation of eigenvectors. In this embodiment, the Sigmoid function is used as the element-wise non-linear activation function.

[0057] The two features are weighted and fused based on the gated vector to obtain the fused feature F:

[0058] (8);

[0059] In the formula, This indicates element-wise multiplication.

[0060] Inputting the fused feature F into a fully connected classification network yields the probability of the vortex presence:

[0061] (9);

[0062] In the formula, The probability of the wake vortex existing. It is a fully connected classification network.

[0063] For the deep learning training process, the data collected by A / B LiDAR was divided into three parts in a 6:2:2 ratio: training data, validation data, and test data. Samples containing wake vortices in each dataset were labeled positive (T), while samples without wake vortices were labeled negative (F). CNN, GRU, and a fusion model of CNN and GRU with complete variables were trained on the training set and validated on the test set. Based on the model parameter design, the GRU model in this invention has two layers, 37 hidden input units in each input direction, and 64 output units. In the CNN model, the stride and kernel size were both set to 3, the training batch size was set to 20, the adaptive Adam optimizer was used, the optimizer momentum factor was set to 0.9, and the learning rate was set to 0.0005. The cross-entropy function was used as the loss function for training the neural network model.

[0064] Furthermore, large-scale datasets are beneficial for neural network model learning. To improve the robustness of the model, this invention trains a deep learning network by applying multiplicative speckle noise enhancement to the training dataset before preprocessing (more closely resembling the speckle characteristics of coherent lidar). This is defined as:

[0065] (10);

[0066] in, and the standard deviation σ from the interval Random sampling is used to form a multi-intensity noise distribution; These are the wind field data values ​​from the original training samples; These are the wind field data values ​​after noise enhancement. If necessary, the enhanced field undergoes physical range trimming and boundary preservation to avoid unreasonable velocity values ​​and structural damage. This noise enhancement, without altering the vortex core center and symmetry, moderately perturbs the radial velocity difference, which helps improve the model's robustness and generalization ability to measurement noise and operating condition disturbances.

[0067] Step S4: Probability and dynamic threshold strategy.

[0068] The probability of the wake at time t is directly used as the model output. As the basis for judgment, an adaptive threshold is generated by combining the on-site atmospheric stability (such as Richardson number or turbulence intensity), background wind speed / direction, aircraft type and operating phase given by runway edge wind sensors and wind profiler radar. Let the eigenvectors be... Taking all the above factors into account, the threshold can be written as:

[0069] (11);

[0070] in, Based on the threshold bias; Generate learnable weight vectors for the model to be used for the threshold; and These are the minimum and maximum values ​​of the threshold, used to limit the threshold range; This is a range clipping function used to restrict input values ​​to [...]. ];

[0071] To suppress transient noise and false triggering, it is required that the recognition results of the two Doppler coherent lidars, A and B, simultaneously satisfy the following for several consecutive frames. The alarm is triggered; repeated events are merged and rate-limited, and are graded as yellow, orange, or red according to the magnitude and duration of exceeding the threshold.

[0072] Step S5: Alarm generation, push notification and logging.

[0073] The results of wake vortex identification in wind field data acquired by lidar using a parallel spatiotemporal deep learning network are as follows: Figure 4 As shown.

[0074] The system integrates aircraft trajectory data acquired from airport ground-based LiDAR / radar, and establishes correlations between wake vortex identification results and specific flight / aircraft type, runway and azimuth, and spatial position / altitude slices to characterize the relationship between wake vortices and aircraft operations. Cross Events are generated only when the duration and dual-radar consistency constraints are met. Each event includes: number, time, spatial location, probability / threshold, level, associated flight / aircraft type, and recommended action. Output channels include: tower display and audio-visual alerts, operating system interface (REST / message queue), and crew-side FMS / EFB messages. Events undergo merging / deduplication, frequency limiting, and lifecycle management. Subsequently, structured messages and visual prompts are pushed to the crew (e.g., FMS / EFB) via the system bus. Message content includes timestamp, runway and bearing, flight / aircraft type, identification probability and threshold, alarm level, uncertainty, spatial location / altitude slice, and suggested action. All events and user interactions are logged, recording identification results, threshold parameters, alarm lifecycle, personnel confirmation, and action results for post-event traceability, statistical evaluation, and model calibration; playback and export are supported when necessary.

[0075] Figure 5The overall architecture of the system 40 provided in this embodiment of the invention is shown. The system consists of five main functional modules, each undertaking different information processing and function execution tasks. System 40 includes: an acquisition module 401, an alignment module 402, an identification module 403, an alarm module 404, and an output module 405.

[0076] This invention also provides an electronic device. Figure 6 A schematic diagram of the structure of an electronic device 50 provided in an embodiment of the present invention includes: a processor 501, a memory 502, and a communication interface 503. Specifically, the program may include program code, which includes computer operation instructions. The memory 502 may include high-speed RAM, and may also include non-volatile memory, such as at least one disk storage device.

[0077] The processor 501 executes computer execution instructions stored in the memory 502 to implement the technical solution of the method embodiment for multi-source data fusion aircraft wake vortex perception and identification described in the foregoing method embodiment. The processor 501 may be a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of the present invention.

[0078] The electronic device 50 may also include a communication interface 503, through which it can communicate and interact with external devices, such as management terminals (e.g., laptops, desktops) or user terminals (e.g., mobile phones, tablets). In specific implementations, if the communication interface 503, memory 502, and processor 501 are implemented independently, then the communication interface 503, memory 502, and processor 501 can be interconnected via a bus to complete communication between them.

[0079] The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc., but this does not mean there is only one bus or only one type of bus. In specific implementations, if the communication interface 503, memory 502, and processor 501 are integrated onto a single chip, then the communication interface 503, memory 502, and processor 501 can communicate through internal interfaces.

[0080] This invention also provides a computer-readable storage medium storing a computer program. When executed by a processor, the computer program implements the technical solution of the above-described aircraft wake vortex real-time identification and warning method embodiment. Its implementation principle and technical effects are similar and will not be repeated here. The computer-readable medium may include random access memory, read-only memory, read-only optical disc or other optical disc storage, disk storage or other magnetic storage devices, or any other medium on which the required program code is intended to be stored or in the form of instructions or data structures, and may be accessed by a computer.

[0081] This invention also provides a computer program product, including a computer program that, when executed by a processor, implements the technical solution of the above-described aircraft wake vortex real-time identification and warning method embodiment. Its implementation principle and technical effects are similar and will not be repeated here. In the specific implementation of its terminal device or server, the processor can be a central processing unit, or other general-purpose processors, digital signal processors, application-specific integrated circuits, etc. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention embodiment can be directly manifested as execution by a hardware processor, or execution by a combination of hardware and software modules in the processor.

[0082] When the technical solution of this invention is implemented in software and provided or used in product form, it can be placed in a storage medium that can be read by a computing device. Based on this, the overall solution of this invention, or its practically contributing parts, can be presented through a software product. This software product includes a computer program or several instructions stored on a corresponding medium to drive a computer device to execute all or part of the steps of the methods described in the embodiments of this invention.

[0083] When integrated units or modules are implemented in hardware, they can take various forms, such as digital circuits and analog circuits. The actual structure of such hardware can utilize devices such as transistors and memristors, but is not limited to these. Unless otherwise specified, the processor can be any suitable hardware processing unit, such as a CPU, GPU, FPGA, DSP, ASIC, etc. Similarly, unless otherwise specified, the storage unit can be any applicable magnetic or magneto-optical storage medium, including but not limited to resistive random access memory (RANM), dynamic random access memory (DRAM), static random access memory (SRAM), enhanced DRAM, high-bandwidth memory, hybrid memory cubes, etc.

[0084] When integrated units or modules are implemented as software programs and sold or deployed as independent products, they can be stored in computer-readable storage media. Based on this, the technical solutions of this invention, especially the improvements to the prior art, or all of its functions, can be embodied in a software product. This software product can be stored in any storage medium and includes instructions for causing computing devices (such as personal computers, servers, network devices, etc.) to execute the method steps described in the embodiments of this invention; these steps can be all or part of the steps.

Claims

1. A method for real-time identification and alarm of aircraft wake vortices through multi-source data fusion, characterized in that, Includes the following steps: Step 1: Multi-source data acquisition: Multi-source data is acquired through airport ground-based LiDAR / radar, runway edge wind sensors, two Doppler coherent lidars (A and B), and wind profiler radar. Step 2: Spatiotemporal alignment and fusion: Perform unified timestamp alignment on the data from the airport ground-based LiDAR / radar, Doppler laser coherent radar, and wind profiler radar to obtain the wake vortex dataset; Step 3: Wake Vortex Identification: Construct a parallel spatiotemporal deep learning wake vortex identification model, including gated recurrent units and convolutional neural networks; The wake vortex dataset is preprocessed into two input parts: sequence input. and block input Input the sequence Input-gated recurrent units extract time-domain features and block inputs. Input is fed into a convolutional neural network to extract spatial domain features; The time-domain features and spatial-domain features are then fused to predict the identification result and existence probability of the wake vortex. Step 4: Probability and Dynamic Threshold Strategy: Combine dynamic probability thresholds for judgment to suppress false alarms and false negatives; Step 5: Alarm generation, push and recording: Generate wake vortex identification results and judgment results, push the output to the tower and unit to each terminal for prompts and record the logs; The probability and dynamic threshold strategy specifically involves directly using the model output to determine the probability of the wake vortex's existence at time t. As the basis for judgment, an adaptive threshold at time t is generated by combining the on-site atmospheric stability, background wind speed / direction, aircraft type, and operational phase provided by the runway edge wind sensor and wind profiler radar. ; Let the eigenvectors be... Incorporating factors such as on-site atmospheric stability, background wind speed / direction, aircraft type, and operational phase, the threshold is written as: ； in, Based on the threshold bias; Generate learnable weight vectors for the model to be used for the threshold; and These are the minimum and maximum values ​​of the threshold, used to limit the threshold range; This is a range clipping function used to restrict input values ​​to [...]. ]; The recognition results of two Doppler coherent lidars, A and B, simultaneously satisfy the following for several consecutive frames. If the threshold is exceeded, an alarm will be triggered; repeated events will be merged and rate-limited, and prompts will be given based on the magnitude and duration of the exceedance.

2. The method for real-time identification and alarm of aircraft wake vortices based on multi-source data fusion according to claim 1, characterized in that, In step 1, the vertical distance between the two Doppler coherent lidars, A and B, does not exceed 10m, and the scanning cycle is half a cycle apart. During the detection process, a real-time wind field dataset is generated by scanning a profile along the direction perpendicular to the aircraft's flight using a range and altitude display scanning mode. The radar measurements include timestamps and position parameters, including distance, azimuth, and elevation. The wind profiler radar simultaneously acquires meteorological elements near the airport and the radar, including wind speed, wind direction, temperature, humidity, boundary layer stability, and turbulence intensity, to obtain background wind field and atmospheric stability data. The three-dimensional trajectory information of arriving and departing aircraft, including latitude / longitude, distance, altitude, and heading / ground speed, is obtained through the airport ground-based radar and monitoring system.

3. The method for real-time identification and alarm of aircraft wake vortices based on multi-source data fusion according to claim 1, characterized in that, Step 1 further includes: using the wind field information acquired by the two Doppler coherent lidars, A and B, as independent input channels, and having a deep learning network automatically learn feature fusion to form complete wind field data, i.e., the wind field data acquired by the radar. ; Step 2 specifically involves: The aircraft tracks acquired by airport-based LiDAR / radar are transformed from the geographic coordinate system to the runway coordinate system with the runway centerline as the reference. Resampling and interpolation are then performed on a unified time step and spatial grid, resulting in spatiotemporally consistent data blocks. Furthermore, considering the influence of the background wind field, the wind field input data from the two Doppler coherent lidars, A and B, are normalized to facilitate training and recognition by the parallel deep learning network. The calculation expression is as follows: ； In the formula, Wind field data acquired by radar, σ and σ' are the mean and standard deviation calculated within a given time window or spatial neighborhood, respectively. This is a standardized result.

4. The method for real-time identification and alarm of aircraft wake vortices based on multi-source data fusion according to claim 3, characterized in that, In step 3, the sequence is input The specific steps for extracting time-domain features from the input gated recurrent unit are as follows: Step 311: The gated loop unit inputs at time t. Compared to the previous hidden state As input, the hidden state at time t is output by updating and resetting the gates to control information forgetting and introducing. ; A bidirectional gated cyclic unit is used to bidirectionally encode and learn the representation of the time-domain information of two Doppler coherent lidars, A and B, thereby focusing on the complete past and future context information of each point in the input sequence. Step 312: Use A layered bidirectional GRU network processes radial velocity wind fields, including forward and backward sequences with multiple individual hidden layers, which are then feedforward and fused into the same output layer. The hidden layer vectors are obtained by learning the characteristics of the forward and backward input processes, respectively. and : ， ； In the formula, For GRU mapping operators; and These represent the standardization results. Forward and backward inputs when used as sequence data; Step 313: Obtain the fused features by concatenating the hidden layer vectors output from the forward and backward input feature processes. The calculation is as follows: ； In the formula, This is a time-domain fusion feature.

5. The method for real-time identification and alarm of aircraft wake vortices based on multi-source data fusion according to claim 4, characterized in that, In step 3, the block input The input convolutional neural network extracts spatial domain features, specifically: Step 321: Divide the wind field data into multiple patches, and then apply a convolutional neural network to the resulting patch set. The convolution process is represented as follows: ； In the formula, To represent the feature map output by the convolutional layer, This represents the set of convolutional kernels with output channels c. It is an element-wise nonlinear activation function; Step 322: After each convolutional layer, activation is applied to generate spatial domain feature maps, thus creating spatial domain feature maps. for: ； In the formula, This is a feature extraction function based on a convolutional neural network. The function includes several convolutional layers, activation layers, and optional normalization layers, used to extract features from the input feature map. Extract spatial domain features from; Step 323: Spatiotemporal Feature Fusion: Using a gated weighted fusion module, the temporal domain features are first fused. Spatial domain characteristics Mapping to a consistent dimension: ； In the formula, and For learnable parameters, , They are respectively , Uniform dimensional features after linear mapping; Step 324: Concatenate the two unified-dimensional features and input them into the gating unit to obtain the gating vector. : ； In the formula, and For learnable parameters, This indicates the concatenation of eigenvectors; Step 325: Based on the gated vector, the two unified-dimensional features are weighted and fused to obtain the fused feature F: ； In the formula, Indicates element-wise multiplication; Step 326: Input the fused feature F into the fully connected classification network to obtain the probability of the vortex presence: ； In the formula, The probability of the wake vortex existing. It is a fully connected classification network.

6. The method for real-time identification and alarm of aircraft wake vortices based on multi-source data fusion according to claim 5, characterized in that, When training a deep learning network, multiplicative speckle noise enhancement is applied to the training dataset before preprocessing, and it is defined as follows: ; in, and the standard deviation σ from the interval Random sampling is used to form a multi-intensity noise distribution; These are the wind field data values ​​from the original training samples; These are the wind field data values ​​after noise enhancement.

7. The method for real-time identification and alarm of aircraft wake vortices based on multi-source data fusion according to claim 6, characterized in that, In step 5, the system integrates aircraft trajectory data acquired by airport ground-based LiDAR / radar, and establishes a correlation between the wake vortex identification results and specific flight / aircraft type, runway and azimuth, as well as spatial position / altitude slice elements, to characterize the relationship between the wake vortex and aircraft operation; when the wake vortex exists, the probability... Crossing the adaptive threshold Events are generated when the duration and dual-radar consistency constraints are met. The event includes: number, time, spatial location, probability / threshold, level, associated flight / aircraft type, and recommended action. The output channels include: tower display and audio-visual alerts, operating system interface, and crew-side FMS / EFB messages; merging / deduplication, frequency limiting, and lifecycle management are performed on events; subsequently, structured messages and visual prompts are pushed to the crew via the system bus. The message content includes timestamps, runway and orientation, flight / aircraft type, probability of existence and threshold, alarm level, uncertainty, spatial location / altitude slices, and suggested actions; all events and user interactions are recorded in log form, which records the identification results, threshold parameters, alarm lifecycle, personnel confirmation, and handling results for post-event traceability, statistical evaluation, and model calibration.

8. A real-time aircraft wake vortex identification and warning system based on multi-source data fusion, characterized in that, include: Acquisition module: Multi-source data acquisition is performed through airport ground-based LiDAR / radar, runway edge wind sensor, two Doppler coherent lidars (A and B), and wind profiler radar; Alignment module: Performs unified timestamp alignment on the data from various sources, including airport ground-based LiDAR / radar, Doppler laser coherent radar, and wind profiler radar, to obtain the wake vortex dataset; Recognition Module: Constructs a parallel spatiotemporal deep learning wake vortex recognition model, including gated recurrent units and convolutional neural networks; The wake vortex dataset is preprocessed into two input parts: sequence input. and block input Input the sequence Input-gated recurrent units extract time-domain features and block inputs. Input is fed into a convolutional neural network to extract spatial domain features; The time-domain features and spatial-domain features are then fused to predict the identification result and existence probability of the wake vortex. Alarm module: Combines dynamic probability thresholds for judgment to suppress false alarms and missed alarms; Output module: alarm generation, push and logging: generates wake vortex identification results and judgment results, pushes them to the tower and unit, outputs them to each terminal for prompts, and records logs; The method of combining dynamic probability thresholds to suppress false alarms and false negatives specifically involves: directly using the model output at time t to determine the probability of the wake vortex's existence. As the basis for judgment, an adaptive threshold at time t is generated by combining the on-site atmospheric stability, background wind speed / direction, aircraft type, and operational phase provided by the runway edge wind sensor and wind profiler radar. ; Let the eigenvectors be... Incorporating factors such as on-site atmospheric stability, background wind speed / direction, aircraft type, and operational phase, the threshold is written as: ； in, Based on the threshold bias; Generate learnable weight vectors for the model to be used for the threshold; and These are the minimum and maximum values ​​of the threshold, used to limit the threshold range; This is a range clipping function used to restrict input values ​​to [...]. ]; The recognition results of two Doppler coherent lidars, A and B, simultaneously satisfy the following for several consecutive frames. If the threshold is exceeded, an alarm will be triggered; repeated events will be merged and rate-limited, and prompts will be given based on the magnitude and duration of the exceedance.

Images