Multi-model hybrid flight trajectory prediction method based on adversarial generation framework

Abstract

The invention discloses a multi-model hybrid flight trajectory prediction method based on an adversarial generation framework, and the method comprises the steps: carrying out the fusion of local feature extraction and global dependence modeling according to the time-space characteristics of flight trajectory data, and combining adversarial training and a dynamic optimization strategy, a hybrid architecture comprising a CNN-Transform generator and an LSTM discriminator and a dynamic weighting multi-objective loss function are introduced, a high-precision and physically reasonable trajectory prediction framework is constructed, and end-to-end prediction from historical observation to future trajectory is realized. The flight path prediction method is obviously superior to a traditional flight path prediction model in prediction precision, physical rationality and calculation efficiency, the dynamic optimization strategy and data driving design of the flight path prediction method provide a high-reliability and low-delay path prediction solution for an air traffic management system, resource optimization and airspace safety efficiency improvement are indirectly supported, and the flight path prediction method has good application prospects. In the future, the physical guidance constraint is combined, and the multi-dimensional collaborative prediction capability is further enhanced.

Description

Technical Field

[0001] The present invention belongs to the technical field of flight trajectory prediction, and particularly relates to the design of a multi-model hybrid flight trajectory prediction method based on an adversarial generation framework. Background Art

[0002] Flight Trajectory Prediction (FTP), as the core technology of Air Traffic Management (ATM), its accuracy directly affects airspace utilization efficiency and flight safety. Existing flight trajectory prediction technologies are mainly divided into the following two categories:

[0003] (1) Physics-based methods: Such methods predict trajectories by establishing aerodynamic models and kinematic equations (such as particle models), which rely on accurate dynamic parameters and environmental assumptions. However, in actual flight scenarios, it is difficult to model dynamic factors such as complex weather disturbances and control instruction coupling effects, resulting in limited generalization ability of the model in real environments.

[0004] (2) Data-driven methods: including traditional machine learning methods, deep learning methods, and Generative Adversarial Networks (GANs). Among them, traditional machine learning methods, such as K-means clustering and Gaussian Mixture Model (GMM), rely on artificial feature design and statistical assumptions, and it is difficult to capture the non-linear relationships of high-dimensional spatio-temporal data. Deep learning methods, such as LSTM / CNN-LSTM, although they can model temporal dependencies, have insufficient local spatio-temporal feature extraction, resulting in lagged prediction of maneuvering actions (such as turning and climbing), and over-reliance on Mean Squared Error (MSE) loss, and the generated trajectories have poor physical rationality; another example is Transformer, which captures long-time domain dependencies through self-attention mechanisms, but its parallel computing characteristics make it less sensitive to local motion patterns. Generative Adversarial Networks (GANs) improve the realism of trajectories through adversarial training, but traditional GANs adopt a single adversarial loss framework, making it difficult to balance geometric accuracy and physical constraints, resulting in limited multi-objective optimization ability.

[0005] The above defects limit the prediction accuracy and engineering applicability of existing methods in complex airspace scenarios, and there is an urgent need for a new framework that integrates multi-scale feature modeling and dynamic optimization strategies. Summary of the Invention

[0006] The purpose of the present invention is to solve the problems of poor physical rationality, insufficient sensitivity to local motion patterns, and limited multi-objective optimization ability existing in existing flight trajectory prediction methods, and propose a multi-model hybrid flight trajectory prediction method based on an adversarial generation framework.

[0007] The technical solution of the present invention is: A multi-model hybrid flight trajectory prediction method based on an adversarial generation framework, including the following steps:

[0008] S1. Obtain the real flight trajectory data and historical flight trajectory data of the aircraft.

[0009] S2. Build a generator G based on CNN and Transformer, and build a discriminator D based on LSTM.

[0010] S3. Conduct adversarial training on the generator G and the discriminator D using the real flight trajectory data and historical flight trajectory data.

[0011] S4. Use the trained generator G for flight trajectory prediction.

[0012] Furthermore, the generator G includes a CNN feature extraction layer, a position encoding module, a Transformer encoding module, and a linear layer projection module connected in sequence.

[0013] Furthermore, the CNN feature extraction layer includes two layers of 1D convolution. The input channel number of the first layer of 1D convolution is 3, the output channel number is 64, the convolution kernel size is 3, and the convolution stride is 1. The input channel number of the second layer of 1D convolution is 64, the output channel number is 128, the convolution kernel size is 3, and the convolution stride is 1. After each layer of 1D convolution, a ReLU activation function, a batch normalization unit BatchNorm, and a regularization unit Dropout are connected in sequence.

[0014] Furthermore, the position encoding module uses sine and cosine functions to perform position encoding on the local spatio-temporal features extracted by the CNN feature extraction layer. The specific formula is:

[0015]

[0016] where PE represents the position encoding result, pos represents the time step index, l represents the feature dimension index, and d represents the dimension of the local spatio-temporal features extracted by the CNN feature extraction layer.

[0017] Furthermore, the Transformer encoding module includes 3 layers of Transformer encoders. Each layer of Transformer encoder includes a multi-head attention mechanism, a feed-forward network FFN, a residual connection, and a layer normalization unit connected in sequence.

[0018] The multi-head attention mechanism includes 4 attention heads. The attention calculation formula for each attention head is:

[0019]

[0020] where Attention(·) represents attention, Q represents the query matrix, K represents the key matrix, V represents the value matrix, d k represents the dimension of a single attention head, and H represents the number of attention heads.

[0021] The feed-forward network FFN includes a linear layer and a ReLU activation function that map 128-dimensional features to 512 dimensions and then back to 128 dimensions.

[0022] Furthermore, the specific formula of the linear layer projection module is:

[0023]

[0024] where represents the predicted flight trajectory data output by the generator G, Linear(128→3)(·) represents the linear layer that maps 128-dimensional features to 3 dimensions, and F global represents the global spatio-temporal features output by the Transformer encoding module.

[0025] Furthermore, the discriminator D includes an LSTM backbone network and a discriminator head. The LSTM backbone network includes 2 layers of bidirectional LSTM with 64 hidden units; the specific formula of the discriminator head is:

[0026] p = Sigmoid(Linear(64→1)(ReLU(Linear(128→64)(h T ))))

[0027] where p represents the authenticity probability output by the discriminator D, Sigmoid(·) represents the Sigmoid activation function, Linear(64→1)(·) represents the linear layer that maps 64-dimensional features to 1 dimension, ReLU(·) represents the ReLU activation function, Linear(128→64)(·) represents the linear layer that maps 128-dimensional features to 64 dimensions, and h T represents the final hidden state output by the LSTM backbone network.

[0028] Furthermore, step S3 includes the following sub-steps:

[0029] S31. Initialize and set the parameters of the generator G and the discriminator D, and use the Adam optimizer to update the parameters.

[0030] S32. Input the historical flight trajectory data into the generator G to generate predicted flight trajectory data.

[0031] S33. Input the real flight trajectory data and the predicted flight trajectory data into the discriminator D, calculate the discriminator loss and backpropagate to update the parameters of the discriminator D, driving the discriminator D to be able to distinguish between real flight trajectory data and predicted flight trajectory data.

[0032] S34. Calculate the generator loss and backpropagate to update the parameters of the generator G, driving the generator G to generate more realistic and geometrically reasonable predicted flight trajectory data to deceive the discriminator D.

[0033] S35. Repeat steps S32 - S34 until the preset number of iterations is reached to obtain the trained generator G and discriminator D.

[0034] Furthermore, the discriminator loss L in step S33 D Specifically:

[0035]

[0036] where E[·] represents the expected value, D(y) represents the output probability of the discriminator D for the real flight trajectory data y, represents the output probability of the discriminator D for the predicted flight trajectory data .

[0037] Furthermore, the generator loss L in step S34 G Specifically:

[0038] L G =(1 - λ1)L adv +λ2L L2

[0039]

[0040] where E[·] represents the expected value, λ1 and λ2 are both balance parameters, L adv represents the adversarial loss, represents the output probability of the discriminator D for the predicted flight trajectory data , represents the i-th predicted flight trajectory data, y i represents the i-th real flight trajectory data, N represents the amount of flight trajectory data, ‖·‖2 represents the L2 norm, L L2 represents the L2 loss.

[0041] The beneficial effects of the present invention are as follows: According to the spatio-temporal characteristics of flight trajectory data, the present invention constructs a high-precision and physically reasonable trajectory prediction framework by integrating local feature extraction and global dependence modeling, combining adversarial training and dynamic optimization strategies, introducing a hybrid architecture including a CNN-Transformer generator and an LSTM discriminator, and a dynamic weighted multi-objective loss function, so as to achieve end-to-end prediction from historical observations to future trajectories. The present invention is significantly superior to traditional flight trajectory prediction models in terms of prediction accuracy, physical rationality, and computational efficiency. Its dynamic optimization strategy and data-driven design provide a highly reliable and low-latency trajectory prediction solution for air traffic management systems, indirectly supporting resource optimization and improvement of airspace safety efficiency. In the future, combined with physical guidance constraints, the multi-dimensional collaborative prediction ability will be further enhanced. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 The figure shows a flowchart of a multi-model hybrid flight trajectory prediction method based on an adversarial generation framework provided by an embodiment of the present invention.

[0043] Figure 2 The figure shows the structural diagrams of the generator and discriminator models provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0044] Now, exemplary embodiments of the present invention will be described in detail with reference to the accompanying drawings. It should be understood that the embodiments shown and described in the drawings are merely exemplary, intended to explain the principles and spirit of the present invention, and not to limit the scope of the present invention.

[0045] An embodiment of the present invention provides a multi-model hybrid flight trajectory prediction method based on an adversarial generation framework, as Figure 1 shown, including the following steps S1 to S4:

[0046] S1. Obtain the real flight trajectory data and historical flight trajectory data of the aircraft.

[0047] In an embodiment of the present invention, 19,021 real flight trajectory data collected by an ADS-B system are used to construct a unified data set after outlier processing, missing value interpolation, and trajectory segmentation (at 10-second intervals). The processed data set includes information on three dimensions: longitude, latitude, and altitude. The front part of the real flight trajectory is used as the historical flight trajectory data.

[0048] S2. Construct a generator G based on CNN and Transformer, and construct a discriminator D based on LSTM.

[0049] In an embodiment of the present invention, as Figure 2As shown, the generator G includes a CNN feature extraction layer, a position encoding module, a Transformer encoding module, and a linear layer projection module connected in sequence.

[0050] Among them, the CNN feature extraction layer includes two layers of 1D convolution. The input channel number of the first layer of 1D convolution is 3 (longitude, latitude, altitude), the output channel number is 64, the convolution kernel size is 3, and the convolution stride is 1; the input channel number of the second layer of 1D convolution is 64, the output channel number is 128, the convolution kernel size is 3, and the convolution stride is 1; after each layer of 1D convolution, a ReLU activation function, a batch normalization unit BatchNorm, and a regularization unit Dropout are connected in sequence. The regularization unit Dropout randomly discards 20% of the neurons to prevent overfitting.

[0051] The position encoding module uses sine and cosine functions to perform position encoding on the local spatio-temporal features extracted by the CNN feature extraction layer, injecting time sequence information into the feature space to avoid the temporal confusion problem of CNN. The specific formula is:

[0052]

[0053] Where PE represents the position encoding result, pos represents the time step index, l represents the feature dimension index, and d represents the dimension of the local spatio-temporal features extracted by the CNN feature extraction layer.

[0054] The Transformer encoding module includes 3 layers of Transformer encoders, and finally outputs the global spatio-temporal feature F global , and each layer of Transformer encoder includes a multi-head attention mechanism, a feed-forward network FFN, a residual connection, and a layer normalization unit connected in sequence.

[0055] The multi-head attention mechanism includes 4 attention heads, and the attention calculation formula for each attention head is:

[0056]

[0057] Where Attention(·) represents attention, Q represents the query matrix, K represents the key matrix, V represents the value matrix, and d k represents the dimension of a single attention head, and H = 4 represents the number of attention heads.

[0058] The feed-forward network FFN includes a linear layer that maps 128-dimensional features to 512 dimensions and then maps back to 128 dimensions, and a ReLU activation function.

[0059] The residual connection and the layer normalization unit are used to stabilize the training process.

[0060] The specific formula of the linear layer projection module is:

[0061]

[0062] Among them represents the predicted flight trajectory data output by the generator G, and Linear(128→3)(·) represents a linear layer that maps 128-dimensional features to 3 dimensions, and F global represents the global spatio-temporal features output by the Transformer encoding module.

[0063] In the embodiment of the present invention, as Figure 2 shown, the discriminator D includes an LSTM backbone network and a discriminator head. The LSTM backbone network includes 2 layers of bidirectional LSTM, the number of hidden units is 64, and finally the final hidden state h T .

[0064] As Figure 2 shown, the specific formula of the discriminator head is:

[0065] p = Sigmoid(Linear(64→1)(ReLU(Linear(128→64)(hT))))

[0066] where p represents the authenticity probability output by the discriminator D, Sigmoid(·) represents the Sigmoid activation function, which limits the output to [0,1], Linear(64→1)(·) represents a linear layer that maps 64-dimensional features to 1 dimension, ReLU(·) represents the ReLU activation function, Linear(128→64)(·) represents a linear layer that maps 128-dimensional features to 64 dimensions, and h T represents the final hidden state output by the LSTM backbone network.

[0067] In the embodiment of the present invention, the input of the generator G is historical flight trajectory data, and the output is predicted flight trajectory data. The input of the discriminator D is real flight trajectory data and the predicted flight trajectory data output by the generator G, and returns the authenticity probability of the input trajectory considered by the discriminator D and provides adversarial feedback.

[0068] S3. Perform adversarial training on the generator G and the discriminator D through real flight trajectory data and historical flight trajectory data.

[0069] Step S3 includes the following sub-steps S31 to S35:

[0070] S31. Initialize and set the parameters of the generator G and the discriminator D, and use the Adam optimizer to update the parameters.

[0071] In the embodiment of the present invention, the learning rate of the Adam optimizer is 5×10 -4, β1 = 0.5, β2 = 0.999.

[0072] S32. Input the historical flight trajectory data into the generator G to generate predicted flight trajectory data.

[0073] S33. Input the real flight trajectory data and the predicted flight trajectory data into the discriminator D, calculate the discriminator loss and backpropagate to update the parameters of the discriminator D, driving the discriminator D to be able to distinguish between real flight trajectory data and predicted flight trajectory data.

[0074] In the embodiment of the present invention, the discriminator D maximizes the probability of real samples and minimizes the probability of generated samples. The discriminator loss L D Specifically:

[0075]

[0076] where E[·] represents the expected value, D(y) represents the output probability of the discriminator D for the real flight trajectory data y, represents the output probability of the discriminator D for the predicted flight trajectory data of.

[0077] S34. Calculate the generator loss and backpropagate to update the parameters of the generator G, driving the generator G to generate more realistic and geometrically reasonable predicted flight trajectory data to deceive the discriminator D.

[0078] In the embodiment of the present invention, the purpose of the generator G is to deceive the discriminator D as much as possible, making the discriminator D judge that the samples generated by the generator G are real. The generator loss L G Specifically:

[0079] L G = (1 - λ1)L adv + λ2L L2

[0080]

[0081] where λ1 and λ2 are both balance parameters, L adv represents the adversarial loss, represents the i-th predicted flight trajectory data, y i represents the i-th real flight trajectory data, N represents the amount of flight trajectory data, ‖·‖2 represents the L2 norm, L L2 represents the L2 loss, which is used to constrain the spatial alignment of the predicted flight trajectory data and the real flight trajectory data, making the predicted flight trajectory data as real as possible.

[0082] S35. Repeat steps S32 - S34 until the preset number of iterations is reached to obtain the trained generator G and discriminator D.

[0083] In the embodiments of the present invention, through the alternating optimization and dynamic adjustment in steps S32 to S34, the generator G and the discriminator D are synergistically improved in the game confrontation.

[0084] S4. Use the trained generator G to predict the flight trajectory.

[0085] In the embodiments of the present invention, after the generator G and the discriminator D are trained, the generator G can extract the features of the newly input flight trajectory and generate a flight trajectory that the discriminator D deems to be true as the high-precision flight trajectory prediction result.

[0086] Those of ordinary skill in the art will realize that the embodiments described herein are to assist the reader in understanding the principles of the present invention, and it should be understood that the protection scope of the present invention is not limited to such specific statements and embodiments. Those of ordinary skill in the art can make various other specific deformations and combinations that do not depart from the essence of the present invention according to these technical revelations disclosed in the present invention, and these deformations and combinations are still within the protection scope of the present invention.

Claims

1. A multi-model hybrid flight trajectory prediction method based on an adversarial generation framework, characterized in that It includes the following steps: S1. Obtain the real flight trajectory data and historical flight trajectory data of the aircraft; S2. Build the generator G based on CNN and Transformer, and build the discriminator D based on LSTM; S3. Conduct adversarial training on the generator G and the discriminator D through the real flight trajectory data and historical flight trajectory data; S4. Use the trained generator G for flight trajectory prediction.

2. The multi-model hybrid flight trajectory prediction method based on an adversarial generation framework according to claim 1, characterized in that The generator G includes a CNN feature extraction layer, a position encoding module, a Transformer encoding module, and a linear layer projection module connected in sequence.

3. The multi-model hybrid flight trajectory prediction method based on the adversarial generation framework according to claim 2, wherein The CNN feature extraction layer includes two layers of 1D convolution. The input channel number of the first layer of 1D convolution is 3, the output channel number is 64, the convolution kernel size is 3, and the convolution stride is 1; the input channel number of the second layer of 1D convolution is 64, the output channel number is 128, the convolution kernel size is 3, and the convolution stride is 1; after each layer of 1D convolution, a ReLU activation function, a batch normalization unit BatchNorm, and a regularization unit Dropout are connected in sequence.

4. The multi-model hybrid flight trajectory prediction method based on an adversarial generation framework according to claim 2, wherein The position encoding module uses sine and cosine functions to perform position encoding on the local spatio-temporal features extracted by the CNN feature extraction layer. The specific formula is: where PE represents the position encoding result, pos represents the time step index, l represents the feature dimension index, and d represents the dimension of the local spatio-temporal features extracted by the CNN feature extraction layer.

5. The multi-model hybrid flight trajectory prediction method based on the adversarial generation framework according to claim 4, wherein, The Transformer encoding module includes 3 layers of Transformer encoders. Each layer of Transformer encoder includes a multi-head attention mechanism, a feed-forward network FFN, a residual connection, and a layer normalization unit connected in sequence; The multi-head attention mechanism includes 4 attention heads. The attention calculation formula for each attention head is: Among them, attention(·) represents attention, Q represents the query matrix, K represents the key matrix, V represents the value matrix, and d k represents the dimension of a single attention head, and H represents the number of attention heads; The feed-forward network FFN includes a linear layer that maps 128-dimensional features to 512 dimensions and then maps back to 128 dimensions, and a ReLU activation function.

6. The multi-model hybrid flight trajectory prediction method based on the adversarial generation framework according to claim 2, wherein, The specific formula of the linear layer projection module is: Among them represents the predicted flight trajectory data output by the generator G, and Linear(128→3)(·) represents a linear layer that maps 128-dimensional features to 3 dimensions, F global represents the global spatio-temporal features output by the Transformer encoding module.

7. The multi-model hybrid flight trajectory prediction method based on the adversarial generation framework according to claim 1, wherein The discriminator D includes an LSTM backbone network and a discriminant head. The LSTM backbone network includes 2 layers of bidirectional LSTM with 64 hidden units; the specific formula of the discriminant head is: Sigmoid = Sigmoid(Linear(64→1)(ReLU(Linear(128→64)(h T )))) where p represents the authenticity probability output by the discriminator D, Sigmoid(·) represents the Sigmoid activation function, Linear(64→1)(·) represents the linear layer that maps 64-dimensional features to 1 dimension, ReLU(·) represents the ReLU activation function, Linear(128→64)(·) represents the linear layer that maps 128-dimensional features to 64 dimensions, and h T represents the final hidden state output by the LSTM backbone network.

8. The multi-model hybrid flight trajectory prediction method based on an adversarial generation framework according to claim 1, characterized in that The step S3 includes the following sub-steps: S31. Initialize and set the parameters of the generator G and the discriminator D, and use the Adam optimizer for parameter update; S32. Input the historical flight trajectory data into the generator G to generate predicted flight trajectory data; S33. Input the real flight trajectory data and the predicted flight trajectory data into the discriminator D, calculate the discriminator loss and backpropagate to update the parameters of the discriminator D, driving the discriminator D to be able to distinguish between real flight trajectory data and predicted flight trajectory data; S34. Calculate the generator loss and backpropagate to update the parameters of the generator G, driving the generator G to generate more realistic and geometrically reasonable predicted flight trajectory data to deceive the discriminator D; S35. Repeat steps S32 - S34 until the preset number of iterations is reached to obtain the trained generator G and discriminator D.

9. The multi-model hybrid flight trajectory prediction method based on an adversarial generation framework according to claim 8, wherein, The discriminator loss L in the step S33 D Specifically: where E[·] represents the expected value, and D(y) represents the output probability of discriminator D for the real flight trajectory data y. represents the output probability of discriminator D for the predicted flight trajectory data of.

10. The multi-model hybrid flight trajectory prediction method based on the adversarial generation framework according to claim 8, wherein The generator loss L in step S34 G Specifically: L G = (1 - λ1)L adv + λ2L L2 Among them, E[·] represents the expected value, both λ1 and λ2 are balance parameters, and L adv represents the adversarial loss, represents the output probability of the discriminator D for the predicted flight trajectory data and represents the i-th predicted flight trajectory data, and y i represents the i-th true flight trajectory data, N represents the amount of flight trajectory data, ‖·‖2 represents the L2 norm, and LL2 represents the L2 loss.

Images