Method for predicting risk of bird colliding with navigation channel

Abstract

The invention provides a bird collision channel risk prediction method, and particularly relates to the technical field of aviation safety, and the method comprises the steps: obtaining bird situation data, airplane flight data, meteorological data and other multi-source data of a preset time window in an airport channel region, and carrying out the preprocessing to obtain standardized data; and then inputting the standardized data into a multi-stream transform model to generate a prediction probability of the bird-hit channel. And according to comparison between the prediction probability and a preset risk level threshold value, dividing a current channel risk level, and triggering a corresponding alarm prevention and control decision based on the risk level, thereby effectively preventing occurrence of bird collision accidents. According to the method, the technical problem of how to effectively improve the accuracy and timeliness of bird strike risk early warning so as to reduce the aviation accident rate is solved, so that real-time accurate prediction and automatic prevention and control decision of the bird strike risk are realized, and the aviation accident risk is remarkably reduced.

Description

Technical Field

[0001] The present invention relates to the field of aviation safety technology, and in particular to a method for predicting bird strike channel risks. Background Art

[0002] With the continued growth of air transport volume and increased efforts to protect the ecological environment, the risk of temporal and spatial overlap between bird activity and flight operations around airports has significantly increased. Traditional bird strike prevention systems rely on single radar monitoring or manual judgment, and are limited by data dimensions (such as ignoring the dynamic impact of meteorological changes on bird behavior and the coupling relationship between aircraft real-time flight paths and bird movements). This results in low risk warning accuracy and delayed prevention and control measures. This makes it difficult to meet the real-time prevention and control needs of bird strike accidents in complex airspace environments, seriously hindering the improvement of aviation safety levels.

[0003] In the existing technology, although the radar-based bird monitoring system (such as patent CN202210002831.7) can count the number and altitude of birds, it does not integrate multi-source data such as meteorological parameters and aircraft dynamic heading, and cannot model cross-modal correlation risks; the Markov chain-based path simulation method (such as patent CN202311297630.5) can predict bird trajectories, but its static model assumptions deviate from the random obstacle avoidance and group behavior of birds in real scenarios, and lack collaborative analysis with the real-time status of flights, resulting in the prediction results being out of touch with the actual risks, making it difficult to support accurate prevention and control decisions.

[0004] To sum up, how to effectively improve the accuracy and timeliness of bird strike risk warnings to reduce the incidence of aviation accidents is a technical problem that urgently needs to be solved. Summary of the Invention

[0005] The main purpose of the present invention is to provide a method for predicting bird strike route risks, so as to at least solve the technical problem of how to effectively improve the accuracy and timeliness of bird strike risk warnings to reduce the incidence of aviation accidents, thereby realizing real-time and accurate prediction of bird strike risks and automated prevention and control decisions, significantly reducing the risk of aviation accidents.

[0006] In order to achieve the above objectives, the present invention provides a method for predicting bird strike channel risk.

[0007] The present invention provides a method for predicting bird strike risk, the method comprising:

[0008] Acquire multi-source data within a preset time window within the airport flight path area, the multi-source data including bird data, aircraft flight data, and meteorological data;

[0009] Preprocessing the multi-source data including the bird data, the aircraft flight data and the meteorological data to obtain standardized data;

[0010] Inputting the standardized data into a multi-stream transformer model for processing to generate a predicted probability of a bird strike path;

[0011] Classify the current waterway risk level according to the comparison result of the predicted probability and the preset risk level threshold;

[0012] Trigger corresponding alarm prevention and control decisions based on the risk level.

[0013] Specifically, the multi-source data of a preset time window in the airport flight path area is obtained, wherein the multi-source data includes bird data, aircraft flight data, and meteorological data, including:

[0014] Collecting the bird data by radar and camera, wherein the bird data includes the flight path, number and speed of the birds;

[0015] Acquiring the aircraft flight data through a flight management system, wherein the aircraft flight data includes the heading, altitude, and position of the aircraft;

[0016] The meteorological data is collected through a weather station, and the meteorological data includes wind speed, wind direction and visibility.

[0017] Specifically, the preprocessing of the multi-source data including the bird data, the aircraft flight data and the meteorological data to obtain standardized data includes:

[0018] Performing wavelet transform denoising and Kalman filter trajectory smoothing on the bird information data to generate first standardized sub-data;

[0019] Normalizing the format of the aircraft flight data and performing interpolation filling of missing values to generate second standardized sub-data;

[0020] Performing Z-score standardization on the meteorological data to generate third standardized sub-data;

[0021] The first normalized sub-data, the second normalized sub-data, and the third normalized sub-data are merged into the normalized data.

[0022] Specifically, the step of inputting the standardized data into a multi-stream transformer model for processing to generate a predicted probability of a bird strike path includes:

[0023] Processing the first normalized sub-data in the normalized data by a bird's mood encoder to generate a first feature vector;

[0024] Processing the second normalized sub-data in the normalized data by a flying encoder to generate a second feature vector;

[0025] Processing the third normalized sub-data in the normalized data by a meteorological encoder to generate a third feature vector;

[0026] fusing the first feature vector, the second feature vector, and the third feature vector into a fused feature vector through a cross-stream attention module;

[0027] The predicted probability is generated by a decoder based on the fused feature vector.

[0028] Specifically, the dividing the current waterway risk level according to the comparison result of the predicted probability and the preset risk level threshold includes:

[0029] When the predicted probability is ≥0.8, it is judged as a major risk; when 0.8>the predicted probability ≥0.6, it is judged as a large risk; when 0.6>the predicted probability ≥0.4, it is judged as a general risk; when the predicted probability is <0.4, it is judged as a low risk.

[0030] Specifically, the cross-stream attention module dynamically allocates the fusion weights of the first eigenvector, the second eigenvector, and the third eigenvector through gating weights, wherein the gating weights are dynamically calculated based on the meteorological data of the current time window.

[0031] The present application provides a method for predicting bird strike channel risks, which achieves accurate risk assessment by integrating multi-source data. The method focuses on the airport channel area and collects bird data, aircraft flight data and meteorological data within a preset time window to ensure the comprehensiveness and timeliness of the information. Subsequently, these multi-source data are preprocessed and converted into standardized data to lay a solid foundation for subsequent analysis. The standardized data is processed by a multi-stream transformer model to generate a predicted probability of bird strike channels and quantify potential risks. Based on the comparison between the predicted probability and the preset risk level threshold, the current channel risk level is clarified. Finally, the corresponding alarm prevention and control decision is automatically triggered based on the risk level, providing strong support for airport safety management and effectively preventing the occurrence of bird strike incidents. This method solves the technical problem of how to effectively improve the accuracy and timeliness of bird strike risk warnings to reduce the incidence of aviation accidents, thereby realizing real-time and accurate prediction of bird strike risks and automated prevention and control decisions, significantly reducing the risk of aviation accidents. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] The accompanying drawings, which constitute part of this application, are intended to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are intended to explain the present invention and do not constitute an undue limitation of the present invention. In the accompanying drawings:

[0033] Figure 1 Schematic diagram of the flow chart of the bird strike channel risk prediction method provided in this application.

[0034] The above drawings illustrate specific embodiments of the present application, which will be described in more detail below. These drawings and the textual description are not intended to limit the scope of the present application in any way, but rather to illustrate the concepts of the present application to those skilled in the art by reference to specific embodiments. DETAILED DESCRIPTION

[0035] To make the purpose, technical solutions, and advantages of this application more clear, the technical solutions in this application will be clearly and completely described below in conjunction with the drawings in this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of them. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0036] The terms "first," "second," "third," "fourth," and so forth (if any) in the description and claims of the present invention and the accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a particular order or sequential sequence. It should be understood that the terms used in this manner are interchangeable where appropriate, such that the embodiments of the present invention described herein can be practiced in sequences other than those illustrated or described herein.

[0037] In the present invention, words such as "exemplary" or "for example" are used to indicate examples, illustrations, or descriptions. Any embodiment or design described in this application as "exemplary" or "for example" should not be construed as being preferred or advantageous over other embodiments or designs. Rather, the use of words such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.

[0038] This application provides a bird strike route risk prediction method that collects multi-source data, including bird activity, aircraft flight data, and meteorological data, within a preset time window within the airport route area. This data is then preprocessed to generate standardized data. This data is then processed using a multi-stream transformer model to generate a predicted probability of a bird strike. By comparing the predicted probability with a preset risk threshold, the route risk level is determined, and corresponding warning and prevention decisions are triggered to ensure flight safety.

[0039] The following specific embodiments describe in detail the technical solution of the present application and how the technical solution of the present application solves the above-mentioned technical problems. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments. The embodiments of the present application will be described below in conjunction with the accompanying drawings.

[0040] Figure 1 The flowchart of the bird strike risk prediction method provided in this application is intended to explain the bird strike risk prediction method in detail, such as Figure 1As shown, the bird strike route risk prediction method provided in this embodiment includes:

[0041] S101: Acquire multi-source data of a preset time window within an airport flight path area, wherein the multi-source data includes bird data, aircraft flight data, and meteorological data;

[0042] Specifically, the multi-source data of a preset time window in the airport flight path area is obtained, wherein the multi-source data includes bird data, aircraft flight data, and meteorological data, including:

[0043] Collecting the bird data by radar and camera, wherein the bird data includes the flight path, number and speed of the birds;

[0044] Acquiring the aircraft flight data through a flight management system, wherein the aircraft flight data includes the heading, altitude, and position of the aircraft;

[0045] The meteorological data is collected through a weather station, and the meteorological data includes wind speed, wind direction and visibility.

[0046] The specific implementation of step S101 is as follows:

[0047] 1. Set data acquisition parameters

[0048] Define the preset time window as 30 minutes, and update the data collection cycle every 5 minutes.

[0049] The airport flight path area is set as a circular airspace with a radius of 20 kilometers centered on the airport runway, and its boundaries are defined by geographic coordinates (for example, longitude range X1-X2, latitude range Y1-Y2).

[0050] 2. Collect bird data

[0051] Radar configuration: Deployed X-band pulse Doppler radar (model: RD-2400), with a radar scanning frequency of 2Hz, a detection range of 30 kilometers, a vertical scanning angle of 0° to 90°, and a horizontal scanning angle of 360° omnidirectional.

[0052] Camera configuration: Install an infrared thermal imaging camera (model: ThermCam-T20) with a resolution of 1920×1080, a frame rate of 30fps, infrared fill light enabled in night mode, and a detection wavelength range of 8-14μm.

[0053] Data processing: Radar data output format: Contains the polar coordinates (range, azimuth, altitude), velocity vector (radial velocity, tangential velocity) and timestamp of the bird flock target.

[0054] Camera data output format: Use the YOLOv5 object detection algorithm to identify bird flocks in video frames and output the flock's bounding box coordinates, population count, and movement direction (calculated as pixel displacement).

[0055] Fusion of radar and camera data: Using the Hungarian algorithm for target matching to generate the bird's flight trajectory (the trajectory point sequence includes longitude, latitude, and altitude. For example, trajectory point P1 = (116.405°E, 39.905°N, 500m)). The Hungarian algorithm is used for target matching to generate the bird's flight trajectory, specifically including:

[0056] (1) Data alignment preparation

[0057] Coordinate system alignment:

[0058] Convert the bird flock targets detected by the radar (polar coordinate system: distance, azimuth, altitude) to the WGS-84 geographic coordinate system (longitude, latitude, altitude) consistent with the camera, using the longitude and latitude of the radar installation location as the reference point during the conversion.

[0059] The pixel coordinates of the bird flock detected by the camera (the center point of the bounding box) are converted to the WGS-84 geographic coordinate system based on the pre-calibrated parameters of the camera (including focal length, installation height, and pitch angle).

[0060] Time alignment: If the timestamp deviation between radar and camera data exceeds 50 milliseconds, linear interpolation is performed on the data within the time deviation to ensure that the radar and camera data at the same time are aligned.

[0061] (2) Constructing the cost matrix of target matching

[0062] a. For each target detected by the radar and each target detected by the camera, calculate the matching cost between the two:

[0063] Spatial distance: the straight-line distance between the two geographic coordinates (unit: meter);

[0064] Velocity difference: the difference between the radial velocity of the radar target and the velocity of the camera target (unit: m / s);

[0065] Comprehensive cost: spatial distance accounts for 70% of the weight, speed difference accounts for 30% of the weight, and the sum is the final cost.

[0066] b. Fill all matching costs of radar and camera targets into the matrix to form an M×N cost matrix (M is the number of radar targets, N is the number of camera targets).

[0067] (3) Perform Hungarian algorithm matching

[0068] a. Input the cost matrix and use the Hungarian algorithm to find the optimal matching pair to minimize the total matching cost.

[0069] b. Algorithm core operations:

[0070] Reduce the rows and columns of the cost matrix (subtract the minimum value of the row or column);

[0071] Use the least number of lines to cover all zero elements. If the number of covering lines is less than the matrix dimension, adjust the matrix and repeat until the optimal match is found.

[0072] c. The output result is a list of matching pairs of radar targets and camera targets. Unmatched targets are marked as invalid data.

[0073] (4) Generate the bird's flight trajectory

[0074] Position fusion: For successfully matched radar and camera targets, the weighted average of their longitude and latitude (radar weight 60%, camera weight 40%) is taken as the fused position.

[0075] Trajectory smoothing: Kalman filtering is applied to the fused trajectory point sequence. During filtering, the state vector is set to include position and velocity, and the observation noise parameter is determined based on the calibration error between the radar and the camera.

[0076] 3. Obtain aircraft flight data

[0077] Data source: ADS-B (Automatic Dependent Surveillance-Broadcast) data interface connected to the airport flight management system (system name: AirTraffic-Control v3.0).

[0078] Data fields:

[0079] (1) Heading: Angle value based on magnetic north (0°-359°), with an accuracy of 0.1°.

[0080] (2) Altitude: barometric altitude value, in meters (e.g. 3000m), with a sampling interval of 10 seconds.

[0081] (3) Location: longitude and latitude in the WGS-84 coordinate system (e.g., 116.408°E, 39.907°N).

[0082] Data synchronization: Synchronize with the UTC time server through the NTP protocol (Network Time Protocol) to ensure that the timestamp error is less than 1 millisecond.

[0083] 4. Collect meteorological data

[0084] Weather station configuration: Three automatic weather stations (model: WS-5000) are deployed, located on the east, west and south sides of the runway, 1 km away from the runway center.

[0085] Sensor parameters:

[0086] (1) Wind speed: Ultrasonic wind speed sensor, range 0-50m / s, accuracy ±0.2m / s.

[0087] (2) Wind direction: electronic compass sensor, range 0°-360°, accuracy ±1°.

[0088] (3) Visibility: Forward scatter visibility meter, range 10m-20km, accuracy ±10%.

[0089] Data transmission: Meteorological data is sent to the central server via the LoRa wireless module, with a data sampling interval of 1 minute.

[0090] This step achieves full-dimensional, high-precision data collection of bird movements, aircraft dynamics, and the meteorological environment in the airport's flight path area by clearly defining the data collection time window, spatial scope, and equipment parameters. Combining multimodal data fusion between radar and cameras, real-time interface calls to the flight management system, and high-frequency sampling from weather stations, this step addresses the problem of missed detections by a single sensor in complex weather conditions. ADS-B data ensures the high timeliness of aircraft locations, while the multi-weather station layout eliminates spatial bias in localized meteorological data, providing a reliable data foundation for subsequent risk prediction.

[0091] S102: Preprocessing the multi-source data including the bird data, the aircraft flight data, and the meteorological data to obtain standardized data;

[0092] Specifically, the preprocessing of the multi-source data including the bird data, the aircraft flight data and the meteorological data to obtain standardized data includes:

[0093] Performing wavelet transform denoising and Kalman filter trajectory smoothing on the bird information data to generate first standardized sub-data;

[0094] Normalizing the format of the aircraft flight data and performing interpolation filling of missing values to generate second standardized sub-data;

[0095] Performing Z-score standardization on the meteorological data to generate third standardized sub-data;

[0096] The first normalized sub-data, the second normalized sub-data, and the third normalized sub-data are merged into the normalized data.

[0097] The specific implementation of step S102 is as follows:

[0098] 1. Preprocessing of bird data

[0099] 1.1 Wavelet transform denoising:

[0100] Algorithm selection: Daubechies 4 wavelet basis function (DB4) is used to denoise the flight trajectory noise in the bird data.

[0101] Decomposition layers: Perform 3-layer wavelet decomposition, and filter out high-frequency coefficients using the hard threshold method. The threshold calculation formula is:

[0102]

[0103] in:

[0104] σ is the standard deviation of noise, estimated by the high-frequency coefficients of the first layer after decomposition;

[0105] N is the number of data points.

[0106] Reconstruct the signal: retain the low-frequency approximation coefficients and reconstruct the denoised trajectory data.

[0107] 1.2 Kalman filter trajectory smoothing:

[0108] State model: define the state vector as [longitude, latitude, altitude, longitude speed, latitude speed, vertical speed] T , the state transfer matrix F is:

[0109]

[0110] Where Δt=5 seconds (data sampling interval).

[0111] Observation model: The observation matrix H is:

[0112]

[0113] Noise parameters: process noise covariance matrix Q = diag(0.1 2 ,0.1 2 ,0.1 2 ,0.5 2 ,0.5 2 ,0.5 2 ), the observation noise covariance matrix R = diag(0.3 2 ,0.3 2 ,1.0 2 )(Unit: latitude and longitude, meters).

[0114] 1.3 Output: Generate the first standardized sub-data, which contains the denoised bird flock trajectory (timestamp, longitude, latitude, altitude, speed).

[0115] 2. Preprocessing of aircraft flight data

[0116] 2.1 Format standardization:

[0117] Timestamp alignment: The UTC timestamps of aircraft flight data are uniformly converted to Unix timestamps (millisecond level) and truncated to 5-second intervals (for example: 12:00:00, 12:00:05).

[0118] Coordinate system conversion: Convert the aircraft's WGS-84 latitude and longitude coordinates to the airport's local Cartesian coordinate system (with the origin at the runway center point). The conversion formula is:

[0119]

[0120] Among them: lon0, lat0 are the longitude and latitude of the center point of the runway;

[0121] R = 6371000 meters (radius of the earth).

[0122] 2.2 Missing value interpolation filling:

[0123] Interpolation method: Linear interpolation is used to fill in the missing aircraft altitude data. For example, if the altitude is missing at t = 10 seconds, and the adjacent data are t = 5 seconds (altitude 3000 meters) and t = 15 seconds (altitude 3100 meters), the fill value is:

[0124]

[0125] 2.3 Output: Generate the second standardized sub-data, including the aircraft heading (0°-359°), Cartesian coordinate position (x, y, z), altitude and timestamp in a unified format.

[0126] 3. Preprocessing of meteorological data

[0127] 3.1Z-score normalization:

[0128] Calculation formula: Standardize wind speed, wind direction, and visibility separately. The formula is: Where: μ is the mean of the historical data of the weather station within a 30-day window; σ is the standard deviation.

[0129] Parameter example: wind speed: μ = 5.2 m / s, σ = 1.8 m / s;

[0130] Visibility: μ=10km, σ=3km.

[0131] 3.2 Output: Generate the third standardized sub-data, including the standardized wind speed, wind direction, visibility and timestamp.

[0132] 4. Data Merge

[0133] The first standardized sub-data (bird situation), the second standardized sub-data (aircraft), and the third standardized sub-data (weather) are aligned by timestamp and merged to form a standardized data table. The fields include: timestamp (Unix milliseconds), bird trajectory (longitude, latitude, altitude, speed), aircraft position (x, y, z), heading, standardized wind speed, standardized wind direction, and standardized visibility.

[0134] This step uses wavelet transform denoising to eliminate environmental noise in the bird data, and Kalman filtering to improve trajectory smoothness and continuity. Coordinate system conversion and linear interpolation of aircraft data ensure spatial alignment and data integrity. Z-score normalization of meteorological data eliminates dimensional differences and makes multi-source data comparable. The resulting standardized data provides spatially aligned, denoised, and dimensionally standardized input for the subsequent multi-stream transformer model, significantly improving the accuracy and reliability of bird strike risk prediction.

[0135] S103: Input the standardized data into the multi-stream transformer model for processing to generate a predicted probability of a bird strike route.

[0136] Specifically, the step of inputting the standardized data into a multi-stream transformer model for processing to generate a predicted probability of a bird strike path includes:

[0137] Processing the first normalized sub-data in the normalized data by a bird's mood encoder to generate a first feature vector;

[0138] Processing the second normalized sub-data in the normalized data by a flying encoder to generate a second feature vector;

[0139] Processing the third normalized sub-data in the normalized data by a meteorological encoder to generate a third feature vector;

[0140] fusing the first eigenvector, the second eigenvector, and the third eigenvector into a fused eigenvector through a cross-stream attention module, wherein the cross-stream attention module dynamically assigns fusion weights of the first eigenvector, the second eigenvector, and the third eigenvector through a gating weight, wherein the gating weight is dynamically calculated based on the meteorological data of the current time window;

[0141] The predicted probability is generated by a decoder based on the fused feature vector.

[0142] The specific implementation of step S103 includes:

[0143] 1. Bird situation encoder processes the first normalized sub-data

[0144] 1.1 Input data: The first standardized sub-data contains the denoised bird flock trajectories (timestamp, longitude, latitude, altitude, speed).

[0145] 1.2 Encoding method: Bidirectional long short-term memory network (Bi-LSTM) is used to extract time series features. The network parameters are as follows:

[0146] Input dimensions: 4 (longitude, latitude, altitude, speed);

[0147] Number of hidden layer nodes: 64;

[0148] Output dimension: 128-dimensional first eigenvector.

[0149] 1.3 Processing Flow:

[0150] Sort the trajectory data by timestamp and input it into Bi-LSTM with every 5 seconds as a time step;

[0151] The hidden state at the last time step is used as the first eigenvector.

[0152] 2. The flying encoder processes the second normalized sub-data

[0153] 2.1 Input data: The second standardized sub-data contains the aircraft's Cartesian coordinates (x, y, z), heading, and altitude.

[0154] 2.2 Encoding method: A fully connected neural network (FCN) is used to extract spatial features. The network structure is as follows:

[0155] Input layer: 5 nodes (x, y, z, heading, height);

[0156] Hidden layer: 2 layers, 32 nodes per layer, activation function is ReLU;

[0157] Output layer: 64-dimensional second eigenvector.

[0158] 3. Meteorological encoder processes the third standardized sub-data

[0159] 3.1 Input data: The third standardized sub-data contains wind speed, wind direction, and visibility after Z-score standardization.

[0160] 3.2 Encoding method: The self-attention mechanism (Transformer Encoder) is used to extract meteorological related features. The parameters are as follows:

[0161] Input dimensions: 3 (wind speed, wind direction, visibility);

[0162] Number of heads: 2 heads for attention;

[0163] Output dimension: 64-dimensional third eigenvector.

[0164] 4. Cross-stream attention module fusion feature vector

[0165] 4.1 Input data: first eigenvector (128 dimensions), second eigenvector (64 dimensions), third eigenvector (64 dimensions).

[0166] 4.2 Gating weight calculation:

[0167] Input the meteorological data (raw wind speed, wind direction, visibility) of the current time window into the gated weight generator:

[0168] The gated weight generator is a fully connected network with a 3D input and a 3D output weight [w1,w2,w3]. The formula is:

[0169] [w1,w2,w3]=Softmax(W·[wind speed,wind direction,visibility] T +b)

[0170] Where W is the weight matrix and b is the bias term.

[0171] 4.3 Feature Fusion

[0172] The three eigenvectors are multiplied by the gated weights and concatenated to generate a 256-dimensional fused eigenvector: fused eigenvector = [w1·first eigenvector, w2·second eigenvector, w3·third eigenvector]

[0173] 5. Decoder generates prediction probabilities

[0174] 5.1 Decoding method: A logistic regression layer is used, which inputs a 256-dimensional fused feature vector and outputs a predicted probability of 0-1.

[0175] Formula: P = σ(W d Fusion feature vector + b d )

[0176] Where: σ is the Sigmoid function; W d is the weight matrix, b d is the bias term.

[0177] 6. Specific structure of the multi-stream transformer model

[0178] 6.1 Model Architecture Parameters:

[0179] 6.1.1 Bird sentiment encoder: uses bidirectional LSTM (Bi-LSTM), input dimensions are 4 (longitude, latitude, altitude, speed), the number of hidden layer nodes is 64, 2 layers are stacked, and the output is a 128-dimensional feature vector.

[0180] Activation function: Tanh is used in the hidden layer, and there is no activation function in the output layer.

[0181] 6.1.2 Flight Encoder: Fully Connected Neural Network (FCN), 5-node input layer (x, y, z, heading, altitude), 2 hidden layers (32 nodes each, ReLU activation), 64-node output layer (no activation function).

[0182] 6.1.3 Meteorological Encoder: Transformer Encoder, input dimension 3 (wind speed, wind direction, visibility), 2 attention heads, feedforward layer dimension 64, output 64-dimensional feature vector.

[0183] 6.1.4 Cross-stream Attention Module:

[0184] The gated weight generator is a single-layer fully connected network that inputs 3D meteorological data and outputs 3D weights, which are normalized by Softmax.

[0185] Feature fusion method: weighted concatenation (weight × feature vector), outputting 256-dimensional fused features.

[0186] 6.1.5 Decoder: Single-layer fully connected network, 256-dimensional input, Sigmoid activation function, output 0-1 probability value.

[0187] 6.2. Model training process

[0188] 6.2.1 Training Dataset:

[0189] Data source: Historical data from 2015 to 2022 for five international airports, including:

[0190] Bird radar / camera records (marking the bird flock's trajectory and whether it causes bird strikes).

[0191] Flight ADS-B data and weather station data at the corresponding time.

[0192] Data division: 70% training set, 15% validation set, and 15% test set.

[0193] Data augmentation: random cropping of time series data and injection of wind speed / visibility noise (±10% perturbation).

[0194] 6.2.1 Training parameters:

[0195] Loss function: Binary Cross-Entropy loss.

[0196] Optimizer: Adam, initial learning rate 0.001, batch size 256.

[0197] Training cycle: Early stopping method (termination when the validation set loss does not decrease for 5 consecutive rounds, maximum 100 rounds).

[0198] Regularization: L2 weight decay (coefficient 1e-4), Dropout rate 0.2 (only FCN hidden layer).

[0199] 6.3. Model performance verification

[0200] Evaluation indicators: test set AUC = 0.93, precision 89.2%, recall 91.5%.

[0201] Comparative experiment: Compared with the single radar model (AUC=0.82) and traditional logistic regression (AUC=0.75), the risk prediction accuracy of this model is improved by 11-18%.

[0202] Ablation experiment: After removing the cross-stream attention module (directly concatenating features), the AUC drops to 0.87, proving the necessity of dynamic weight fusion.

[0203] This step uses a bird behavior encoder to extract the dynamic temporal characteristics of the flock, a flight encoder to capture the spatial state of the aircraft, and a weather encoder to explore the correlations between meteorological conditions. A cross-stream attention module then dynamically assigns weights (for example, increasing the weight of meteorological data in windy weather), ultimately decoding and generating predicted probabilities. This step addresses the fusion challenges caused by the heterogeneity of multimodal data, enabling collaborative modeling of bird flocks, aircraft, and weather conditions, significantly improving the accuracy and real-time performance of risk prediction.

[0204] S104: Classifying the current waterway risk level according to the comparison result between the predicted probability and the preset risk level threshold;

[0205] Specifically, the dividing the current waterway risk level according to the comparison result of the predicted probability and the preset risk level threshold includes:

[0206] When the predicted probability is ≥0.8, it is judged as a major risk; when 0.8>the predicted probability ≥0.6, it is judged as a large risk; when 0.6>the predicted probability ≥0.4, it is judged as a general risk; when the predicted probability is <0.4, it is judged as a low risk.

[0207] The specific implementation includes:

[0208] 1. Definition of preset risk level threshold

[0209] Major risk threshold: predicted probability P ≥ 0.8;

[0210] Greater risk threshold: predicted probability 0.8>P≥0.6;

[0211] General risk threshold: predicted probability 0.6>P≥0.4;

[0212] Low risk threshold: predicted probability P < 0.4.

[0213] 2. Risk Level Determination Process

[0214] 2.1 Input data: the predicted probability P output in step S103 (range 0-1, representing the probability of a bird strike event);

[0215] 2.2 Decision logic:

[0216] If P ≥ 0.8, a major risk level is triggered;

[0217] If 0.8>P≥0.6, a higher risk level is triggered;

[0218] If 0.6>P≥0.4, the general risk level is triggered;

[0219] If P<0.4, a low risk level is triggered.

[0220] 2.3 Output result: risk level label of the current waterway (major risk, high risk, general risk, low risk).

[0221] 3. Explanation of threshold setting basis

[0222] Threshold source: A classification model trained based on historical bird strike accident data. The threshold division is determined by the ROC curve to ensure that the high risk level (P ≥ 0.8) corresponds to samples with an actual accident rate exceeding 90%.

[0223] This step discretizes the continuous predicted probability into four actionable risk levels by clearly defining the mapping relationship between predicted probability and risk level (e.g., 0.8 is the threshold for significant risk). This resolves the issue of ambiguous risk descriptions in traditional methods. Specifically, the statistically validated threshold setting (e.g., significant risk corresponds to a high actual accident rate) ensures the scientific and reliable nature of risk grading, providing a precise basis for subsequent warning and prevention decisions, and significantly improving the efficiency of aviation safety management.

[0224] S105: Triggering corresponding alarm prevention and control decisions based on the risk level.

[0225] This step is implemented in the following situations:

[0226] 1. Major risk (predicted probability ≥ 0.8):

[0227] Start the drone bird-repellent formation and control the drone group to perform coordinated bird-repellent tasks around the waterway. At the same time, activate the high-power laser bird-repellent equipment to perform multi-angle scanning and repelling.

[0228] 2. High risk (0.8 > predicted probability ≥ 0.6):

[0229] Dynamically adjust the output power of laser bird repellent equipment and generate flight detour route suggestions based on real-time meteorological data.

[0230] 3. General risk (0.6 > predicted probability ≥ 0.4):

[0231] Activate the fixed sonic bird repellent device, play preset bird deterrent sound waves, and increase the monitoring frequency of radar and cameras to 5 times per second.

[0232] 4. Low risk (predicted probability < 0.4):

[0233] Record current bird data and risk levels to the airport safety database, and display a low-risk warning sign on the air traffic control interface.

[0234] This step automatically triggers differentiated prevention and control actions based on risk level: High-risk scenarios involve proactively repelling bird flocks through drone formations and laser equipment; medium-risk scenarios dynamically adjust prevention and control intensity and flight paths; and low-risk scenarios prioritize data recording and early warning. By integrating multi-source data with prevention and control equipment, we achieve comprehensive coverage, from proactive intervention to reactive monitoring, ensuring precise alignment of prevention and control measures with risk levels, reducing false alarms and missed alerts, and improving aviation safety response efficiency.

[0235] This embodiment provides a method for predicting the risk of bird strikes on airways. The method focuses on the airport airway area and comprehensively collects multi-source information such as bird data, aircraft flight data and meteorological data within a preset time window to ensure the comprehensiveness and timeliness of risk assessment. By pre-processing these multi-source data and converting them into standardized data, a solid foundation is laid for subsequent analysis. The standardized data is processed using a multi-stream transformer model to generate a predicted probability of bird strikes on airways, thereby achieving a quantitative assessment of the risk. Based on the comparison between the predicted probability and the preset risk level threshold, the risk level of the current airway can be quickly divided. Ultimately, the corresponding alarm prevention and control decision is automatically triggered based on the risk level, providing the airport management department with a timely and effective risk response plan, thereby significantly reducing the impact of bird strikes on flight safety.

[0236] Those skilled in the art will readily appreciate other embodiments of the present application after considering the specification and practicing the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of the present application that follow the general principles of the present application and include common knowledge or customary techniques in the art not disclosed herein. The description and examples are to be considered as exemplary only, and the true scope and spirit of the present application are indicated by the following claims.

[0237] It should be understood that the present application is not limited to the exact structure described above and shown in the drawings, and that various modifications and changes may be made without departing from the scope thereof. The scope of the present application is limited only by the appended claims.

Claims

1. A method for predicting bird strike risk, characterized in that: The method comprises: Acquire multi-source data within a preset time window within the airport flight path area, the multi-source data including bird data, aircraft flight data, and meteorological data; Preprocessing the multi-source data including the bird data, the aircraft flight data and the meteorological data to obtain standardized data; Inputting the standardized data into a multi-stream transformer model for processing to generate a predicted probability of a bird strike path; Classify the current waterway risk level according to the comparison result of the predicted probability and the preset risk level threshold; Trigger corresponding alarm prevention and control decisions based on the risk level.

2. The method according to claim 1, characterized in that The multi-source data of a preset time window in the airport flight path area is obtained, wherein the multi-source data includes bird data, aircraft flight data and meteorological data, including: Collecting the bird data by radar and camera, wherein the bird data includes the flight path, number and speed of the birds; Acquiring the aircraft flight data through a flight management system, wherein the aircraft flight data includes the heading, altitude, and position of the aircraft; The meteorological data is collected through a weather station, and the meteorological data includes wind speed, wind direction and visibility.

3. The method according to claim 1, characterized in that The preprocessing of the multi-source data including the bird data, the aircraft flight data and the meteorological data to obtain standardized data includes: Performing wavelet transform denoising and Kalman filter trajectory smoothing on the bird information data to generate first standardized sub-data; Normalizing the format of the aircraft flight data and performing interpolation filling of missing values to generate second standardized sub-data; Performing Z-score standardization on the meteorological data to generate third standardized sub-data; The first normalized sub-data, the second normalized sub-data, and the third normalized sub-data are merged into the normalized data.

4. The method according to claim 3, characterized in that Inputting the standardized data into a multi-stream transformer model for processing to generate a predicted probability of a bird strike path includes: Processing the first normalized sub-data in the normalized data by a bird's mood encoder to generate a first feature vector; Processing the second normalized sub-data in the normalized data by a flying encoder to generate a second feature vector; Processing the third normalized sub-data in the normalized data by a meteorological encoder to generate a third feature vector; fusing the first feature vector, the second feature vector, and the third feature vector into a fused feature vector through a cross-stream attention module; The predicted probability is generated by a decoder based on the fused feature vector.

5. The method according to claim 1, wherein The step of classifying the current waterway risk level according to the comparison result between the predicted probability and the preset risk level threshold includes: When the predicted probability is ≥0.8, it is judged as a major risk; when 0.8>the predicted probability ≥0.6, it is judged as a large risk; when 0.6>the predicted probability ≥0.4, it is judged as a general risk; when the predicted probability is <0.4, it is judged as a low risk.

6. The method according to claim 4, characterized in that The cross-stream attention module dynamically allocates the fusion weights of the first eigenvector, the second eigenvector and the third eigenvector through gating weights, wherein the gating weights are dynamically calculated based on the meteorological data of the current time window.

Images