A multi-channel-frequency-based flight trajectory prediction method

Abstract

The application provides a kind of flight trajectory prediction method based on multi-channel-frequency, it is related to flight trajectory prediction technical field.The method includes obtaining trajectory sequence data, extracting the time mode of its multiple features and fusing, embedding, obtaining trajectory sequence feature embedding;Based on trajectory sequence feature embedding, extract low-frequency, high-frequency information;Low-frequency, high-frequency information is input into frequency feature extraction module respectively, and low-frequency, high-frequency key information is extracted;Low-frequency, high-frequency key information is converted into low-frequency, high-frequency signal;Based on low-frequency, high-frequency signal, obtain time domain trajectory sequence, and intercept the last point as the predicted trajectory point;Iterative calculation, obtain multiple predicted trajectory points and connect in turn, complete flight trajectory prediction.The method can effectively enhance the perception ability to flight state change, improve the accuracy of flight trajectory prediction, and guarantee the spatiotemporal coherence of trajectory;In general, flight trajectory prediction under more scene types and more complex flight states can be realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of flight trajectory prediction, in particular to a flight trajectory prediction method based on multi-channel-frequency. BACKGROUND

[0002] In air traffic management, flight trajectory prediction is an important content. By providing accurate flight trajectory information to air traffic controllers, airspace conditions can be better monitored, potential conflicts can be detected in time, and the possibility of flight delays can be predicted in advance. In actual aviation operations, accurate and effective flight trajectory prediction is crucial to ensure flight safety and optimize control operations. Through the multi-channel-frequency flight trajectory prediction method, the flight trajectory prediction accuracy can be improved, and the basis for airlines to develop more reasonable flight plans can be provided.

[0003] Existing flight trajectory prediction methods can be divided into three categories, namely dynamic modeling method, state estimation method and machine learning method. The dynamic modeling method establishes kinematic equations to simulate flight patterns and predict future motion trajectories. The state estimation method propagates and estimates the subsequent motion state according to the aircraft properties (such as position, speed, flight speed, etc.). Traditional machine learning methods use a large amount of historical trajectory data to learn the basic patterns and features of aircraft motion, but require expert manual feature design. With the development of deep learning, deep neural networks have gradually emerged, which can automatically learn complex hierarchical feature representations from input data without human feature engineering. The existing deep learning prediction methods have the following problems: first, the model ignores the flight trajectory trend, making it difficult to capture changes in flight state; second, it ignores the time pattern of each variable, making it difficult to capture the correlation between flight variables SUMMARY

[0004] To solve the above technical problems existing in the prior art, the present application aims to provide a flight trajectory prediction method that captures the overall trend of flight and changes in flight motion details, and integrates the time-dependent relationship in flight sequence data.

[0005] Specifically, the present application provides a flight trajectory prediction method based on multi-channel-frequency, and the technical solution specifically includes the following steps:

[0006] Step S1: Obtain trajectory sequence data of a single flight including multiple features; wherein each feature corresponds to a channel;

[0007] Step S2: Extract the time pattern of each feature of the trajectory sequence data through the MLP (Multilayer Perceptron, Multi-Layer Perceptron) network and integrate it to obtain the trajectory sequence feature representation, and then perform embedded expression to obtain the trajectory sequence feature embedding;

[0008] Step S3: embedding the trajectory sequence feature as input, extracting low-frequency information through a first convolutional neural network, and extracting high-frequency information through a second convolutional neural network;

[0009] Step S4: inputting the low-frequency information and the high-frequency information into a frequency feature extraction module respectively to extract low-frequency key information and high-frequency key information; wherein the frequency feature extraction module includes three independent Mamba blocks;

[0010] Step S5: converting the low-frequency key information into a low-frequency signal and converting the high-frequency key information into a high-frequency signal;

[0011] Step S6: taking the low-frequency signal and the high-frequency signal as input, obtaining a time-domain trajectory sequence of the trajectory sequence data through inverse wavelet transform, and intercepting the last point of the time-domain trajectory sequence as a predicted trajectory point;

[0012] Step S7: based on the trajectory sequence data and the predicted trajectory point, iteratively calculating a new predicted trajectory point; connecting all the predicted trajectory points in sequence as the result of flight trajectory prediction.

[0013] Further, before step S1, there is also:

[0014] obtaining a trajectory data set and deleting missing and repeated trajectory data; wherein the trajectory data set includes trajectory sequence data of multiple flights.

[0015] Further, after step S1, there is also:

[0016] unifying the time granularity of the trajectory sequence data through piecewise linear interpolation and resampling;

[0017] cutting the trajectory sequence data of different lengths into the same size of slices X 1:t through a sliding window operation; wherein t is the slice length.

[0018] Preferably, step S2 specifically includes:

[0019] inputting each slice X 1:t in sequence into a fully connected layer and a ReLU activation layer after transposition to obtain the corresponding time pattern D,

[0020] D=ReLu(Linear(X 1:t ) T );

[0021] In the formula, Linear(·) represents a fully connected operation;

[0022] adding the time pattern D transposed to the slice X 1:t to obtain the trajectory sequence feature representation D',

[0023] D′=DT +X 1:t ;

[0024] embedding the trajectory sequence feature representation D' by an Embedding method,

[0025] E = Embedding(D');

[0026] wherein E is the trajectory sequence feature embedding.

[0027] Preferably, the step S3 specifically comprises:

[0028] E low = Conv L (E);

[0029] E high = Conv H (E);

[0030] wherein Conv L (·) and Conv H (·) represent the first convolutional neural network and the second convolutional neural network respectively, E low and E high are the low-frequency information and the high-frequency information respectively.

[0031] Preferably, the step S4 specifically comprises:

[0032] inputting the low-frequency information E low into three independent Mamba blocks respectively, and extracting the low-frequency key information A low through an attention mechanism,

[0033] Q low = Mamba0(E low );

[0034] K low = Mamba1(E low );

[0035] V low = Mamba2(E low );

[0036] A low = Q low ⊙K low ⊙V low ;

[0037] wherein Q low , K low and V low are intermediate calculation quantities; Mamba0(·), Mamba1(·) and Mamba2(·) represent the three independent Mamba blocks respectively;

[0038] high-frequency information E high are input into three independent Mamba blocks respectively, and high-frequency key information A high is extracted through an attention mechanism

[0039] Q high = Mamba3(E high );

[0040] K high = Mamba4(E high );

[0041] V high = Mamba5(E high );

[0042] A high = Q high ⊙K high ⊙V high ;

[0043] wherein Q high , K high and V high are intermediate calculation quantities; Mamba3(·), Mamba4(·) and Mamba5(·) represent three independent Mamba blocks respectively.

[0044] Preferably, step S4 further comprises:

[0045] Mamba0(·) and Mamba3(·) are the same Mamba block.

[0046] Preferably, step S5 specifically comprises:

[0047] the low-frequency key information A low is sequentially subjected to convolution operation and linear processing to obtain a low-frequency signal F low ,

[0048] F low = Linear(Conv l (A low ));

[0049] wherein Conv l (·) represents a third convolution layer;

[0050] the high-frequency key information A high is sequentially subjected to convolution operation and linear processing to obtain a high-frequency signal F high ,

[0051] F high = Linear(Conv h (Ahigh

[0052] Conv h (·) represents the fourth convolutional layer.

[0053] Preferably, the step S6 specifically comprises:

[0054] Y 1:t+1 = IDWT(F low ,F high )

[0055] Y 1:t+1 represents a time-domain trajectory sequence, and IDWT(·) represents an inverse wavelet transform.

[0056] The last point Y t+1 is intercepted as a trajectory prediction point P.

[0057] The technical solution provided by the present application can effectively enhance the perception ability of flight state changes and improve the accuracy of flight trajectory prediction. Through the inverse wavelet transform, the multi-frequency characteristics are converted into a time-domain trajectory sequence, and the continuous trajectory point sequence is generated by combining the iterative prediction, which can guarantee the spatio-temporal coherence of the trajectory. In general, the technical solution provided by the present application can realize flight trajectory prediction in more scene types and more complex flight states. BRIEF DESCRIPTION OF DRAWINGS

[0058] Figure 1 Fig. 1 is a schematic diagram of the overall model structure of the flight trajectory prediction based on multi-channel-frequency in the present application.

[0059] Figure 2 Fig. 4 is a detailed architecture diagram of the Mamba module in the present application.

[0060] Figure 3 Fig. 5 is a comparison diagram of the effect of the flight trajectory prediction provided by the present application and the existing prediction method in the ascending stage of the real data set.

[0061] Figure 4 Fig. 6 is a comparison diagram of the effect of the flight trajectory prediction provided by the present application and the existing prediction method in the descending stage of the real data set.

[0062] Figure 5 Fig. 7 is a comparison diagram of the effect of the flight trajectory prediction provided by the present application and the existing prediction method in the cruising stage of the real data set.

[0063] Figure 6 Fig. 8 is a comparison diagram of the effect of the flight trajectory prediction provided by the present application and the existing prediction method in the turning stage of the real data set. DETAILED DESCRIPTION

[0064] ​Hereinafter, the technical solutions provided by the present application will be further described in detail with reference to the drawings.

[0065] As shown in the figure, the flight trajectory prediction according to the collected trajectory data set comprises the following steps: Figure 1

[0066] Step S1: data preprocessing is performed on the trajectory sequence data;

[0067] Step S2: an MLP network is used to extract and fuse the time patterns of each channel;

[0068] Step S3: two convolutional neural networks are respectively used to preliminarily extract low-frequency information E low and high-frequency information E high ;

[0069] Step S4: a frequency feature extraction module based on Mamba is used to further extract low-frequency key information A low and high-frequency key information A high ;

[0070] Step S5: a convolutional neural network is used to convert the high-frequency and low-frequency key information into low-frequency signal F low and high-frequency signal F high ;

[0071] Step S6: an inverse wavelet transform is used to convert the frequency signals F low and F high into time-domain trajectory sequences, and the last point is intercepted as a predicted trajectory point.

[0072] Step S1 specifically comprises:

[0073] Step S11: the trajectory data set is acquired, and a de-duplication operation is performed, the trajectory sequence data is extracted individually according to the flight number, and the data integrity is ensured; wherein the trajectory sequence data comprises multiple features such as longitude, latitude, height and speed, and each feature corresponds to a channel;

[0074] Step S12: the trajectory data is subjected to piecewise linear interpolation and resampling operation at an interval of 10s, and the time granularity is unified;

[0075] Step S13: the trajectory is sliced and divided through a sliding window operation, and trajectory sequences of different lengths are cut into slices X 1:t of the same size, wherein t is the slice length.

[0076] Step S2 specifically comprises:

[0077] Step S21: the trajectory input slice X 1:t ​The transpose operation is performed, and the time mode D of different features in the input sequence is extracted through the full connection layer and the ReLU activation layer in turn:

[0078] D = ReLu(Linear(X 1:t ) T );

[0079] Step S22: The trajectory sequence slice time mode D is transposed and added to the original input trajectory sequence slice X 1:t , the original channel time mode is retained and fused, and the trajectory sequence feature representation D' is obtained:

[0080] D' = D T + X 1:t ;

[0081] Step S3 specifically includes:

[0082] Step S31: The trajectory sequence slice D' after fusing the channel time mode is embedded into 64 dimensions using the Embedding method to obtain semantic representation E:

[0083] E = Embedding(D');

[0084] Step S32: Embedding E is respectively passed through two convolutional neural networks with different receptive field sizes, and the convolution Conv L focuses on local areas to extract basic mode features E low in the trajectory that evolve slowly, the convolution Conv H with a kernel size of 3x3 and a stride of 1 and padding of 1 is used to capture the relative long-distance dependencies in the trajectory sequence and extract detailed features E high that change rapidly:

[0085] E low = Conv L (E);

[0086] E high = Conv H (E);

[0087] Step S4 specifically includes:

[0088] Step S41: The low-frequency information E low and the high-frequency information E high are respectively passed through three Mamba blocks. Combined with the attention mechanism, three different Mamba blocks are respectively used as query (Q), key (K) and value (V) components to extract their respective key information. Finally, the respective Q, K, V Hadamard product is applied on the low-frequency and high-frequency branches respectively to extract more fine-grained local features A lowand richer context information A high :

[0089] Q = Mamba0(L);

[0090] K = Mamba1(L);

[0091] V = Mamba2(L);

[0092] A = Q O K O V;

[0093] where L = E low or E high , A = A low or A high , Q, K, V are only used to generate attention matrix to extract context key information A, and O represents Hadamard multiplication.

[0094] The specific details of the Mamba block are shown in Figure 2 . Wherein, X t represents the input, Y t represents the output, A1 represents the structured state matrix, B t , C t both represent learnable parameters, and Δt represents the step size for discretizing continuous sequences.

[0095] Step S42: Strengthen the interaction between high-frequency information and low-frequency information by sharing the Mamba0 module operation unit of high-frequency information and low-frequency information:

[0096] Mamba0 = SharedMamba;

[0097] Step S5: The extracted low-frequency key information A low and high-frequency key information A low are respectively sequentially subjected to independent convolution layers and linear layers, to obtain low-frequency signal F low and high-frequency signal F high :

[0098] F low = Linear(Conv l (A low ));

[0099] F high = Linear(Conv h (A high ));

[0100] Wherein, the convolution kernel size of Conv l and Conv h is 3x3, the step is 2, and the padding is 1;

[0101] Step S6 specifically includes:

[0102] Step S61: Use inverse wavelet transform to convert the frequency signal F low and F high Converted into a time-domain trajectory sequence Y 1:t+1 :

[0103] Y 1:t+1 =IDWT(F low ,F high );

[0104] Step S62: Extract the last point as the trajectory prediction point P:

[0105] P = Y t+1 ;

[0106] Step S63: Using the predicted point P as the trajectory node for the next time step, and through multiple iterations of prediction, form a complete predicted trajectory sequence T. L :

[0107] T L ={Y t+1 ,Y t+2 ,…,Y t+n}

[0108] Where n is the length of the predicted trajectory sequence.

[0109] like Figure 3 As shown, the X, Y, and Z axes in the three-dimensional coordinate system represent longitude, latitude, and altitude, respectively. The legend in the upper left corner shows, in order: Input sequence, Ground Truth prediction sequence, WTFTP model prediction sequence, the prediction sequence of this invention's model, LSTM (Long Short-Term Memory) model prediction sequence, and CNNLSTM (Convolutional Neural Networks Long Short-Term Memory) model prediction sequence. Real-world flight datasets were used for model training and validation, and the prediction performance of each model at different spatial locations was demonstrated and evaluated. For example... Figure 3 , 4 As shown in Figures 5 and 6, the experimental results show that the prediction method of this invention exhibits higher accuracy and continuity in different flight states, whether it is the ascent phase, descent phase, cruise phase, or turning phase, and has significant advantages over traditional methods.

[0110] The application not only solves the core problem in flight trajectory prediction, but also significantly improves the accuracy and continuity of the model, and is suitable for trajectory prediction under complex flight states. In view of the key problems of the existing flight trajectory prediction technology, a prediction method based on multi-frequency and multi-channel fusion is proposed, and the technical effect is realized through targeted technical means. First, the traditional model often ignores the overall trend of the flight trajectory, and it is difficult to capture the change of the flight state. The low-frequency and high-frequency characteristics in the trajectory data are decomposed and modeled by the multi-frequency Mamba module, the low-frequency characteristics are used to capture the overall trend of the flight trajectory, and the high-frequency characteristics are used to identify the rapid change of the flight details, so as to effectively enhance the perception ability of the model to the change of the flight state and improve the accuracy of the trajectory prediction. In addition, the existing model often ignores the time mode of each variable, which makes it difficult to capture the correlation between the flight variables. The time mode of each channel is extracted and fused by the MLP network, and the time dependence between the variables is comprehensively identified. Finally, the traditional prediction method has defects in the space-time continuity of trajectory prediction, which affects the overall prediction effect. The multi-frequency characteristics are converted into time domain trajectory sequences by inverse wavelet transform, and combined with iterative prediction to generate continuous trajectory point sequences, so as to ensure the space-time continuity of the trajectory.

[0111] As can be seen from the above, the technical scheme provided by the application can effectively enhance the perception ability to the change of the flight state, and improve the accuracy of the flight trajectory prediction. Through inverse wavelet transform, the multi-frequency characteristics are converted into time domain trajectory sequences, combined with iterative prediction to generate continuous trajectory point sequences, which can ensure the space-time continuity of the trajectory. In summary, the technical scheme provided by the application can realize flight trajectory prediction under more scene types and more complex flight states.

[0112] Further, on the basis of achieving the above beneficial effects, each preferred scheme also achieves the following beneficial effects: deleting missing and repeated trajectory data helps to improve the accuracy and precision of flight trajectory prediction; unifying the time granularity and slicing the trajectory sequence data is beneficial to improve the accuracy and efficiency of flight trajectory prediction.

Claims

1. A flight trajectory prediction method based on multi-channel-frequency, characterized in that, Includes the following steps: Step S1: Obtain trajectory sequence data of a single flight, including multiple features; where each feature corresponds to a channel; Step S2: Extract the temporal pattern of each feature of the trajectory sequence data through the MLP network and fuse them to obtain the trajectory sequence feature representation, and then perform embedded expression to obtain the trajectory sequence feature embedding; Step S3: Using the trajectory sequence feature embedding as input, extract low-frequency information through the first convolutional neural network and extract high-frequency information through the second convolutional neural network; Step S4: Input the low-frequency information and high-frequency information into the frequency feature extraction module respectively to extract the low-frequency key information and high-frequency key information; wherein, the frequency feature extraction module includes three independent Mamba blocks; Step S5: Convert low-frequency key information into low-frequency signals and high-frequency key information into high-frequency signals; Step S6: Using low-frequency and high-frequency signals as input, obtain the time-domain trajectory sequence of the trajectory sequence data through inverse wavelet transform, and extract the last point of the time-domain trajectory sequence as the predicted trajectory point; Step S7: Based on the trajectory sequence data and the predicted trajectory points, iteratively calculate new predicted trajectory points; connect all the predicted trajectory points in sequence to obtain the result of the flight trajectory prediction.

2. The flight trajectory prediction method based on multi-channel frequency as described in claim 1, characterized in that, Step S1 is preceded by: Obtain the trajectory dataset and remove missing and duplicate trajectory data; the trajectory dataset includes trajectory sequence data of multiple flights.

3. The flight trajectory prediction method based on multi-channel frequency as described in claim 1, characterized in that, Step S1 is followed by: By using piecewise linear interpolation and resampling, the temporal granularity of trajectory sequence data is unified; The sliding window operation is used to divide trajectory sequence data of different lengths into slices of the same size X. 1:t Where t is the slice length.

4. The flight trajectory prediction method based on multi-channel-frequency as described in claim 3, characterized in that, Step S2 specifically includes: Each slice X1:t is transposed and then sequentially fed into a fully connected layer and a ReLU activation layer to obtain the corresponding temporal pattern D. D) ReRead(Linear(X 1：t ) T ); In the formula, Linear(·) represents a fully connected operation; Transpose the time pattern D and then compare it with slice X. 1：t Add them together to obtain the trajectory sequence feature representation D′. D′=D T +X 1：t ; The trajectory sequence feature representation D′ is embedded using the Embedding method. E = Embedding(D′); In the formula, E represents the trajectory sequence feature embedding.

5. The flight trajectory prediction method based on multi-channel-frequency as described in claim 4, characterized in that, Step S3 specifically includes: AND low =Conv L (AND); AND high =Conv H (AND); In the formula, Conv L (·) and Conv H (·) represent the first convolutional neural network and the second convolutional neural network, respectively. low and E high These are low-frequency information and high-frequency information, respectively.

6. The flight trajectory prediction method based on multi-channel frequency as described in claim 5, characterized in that, Step S4 specifically includes: Low-frequency information E low Three independent Mamba blocks are input respectively, and low-frequency key information A is extracted using an attention mechanism. low , Q low =Mamba0(E low ); K low =Mamba1(E low ); V low =Mamba2(E low ); A low =Q low ⊙K low ⊙V low ; In the formula, Q low K low and V low All are intermediate computational quantities; Mamba0(·), Mamba1(·), and Mamba2(·) represent three independent Mamba blocks, and ⊙ represents Hadamard multiplication; High-frequency information E high Three independent Mamba blocks are input respectively, and high-frequency key information A is extracted using an attention mechanism. high , Q high =Mamba3(E high ); K high =Mamba4(E high ); V high =Mamba5(E high ); A high =Q high ⊙K high ⊙V high ; In the formula, Q high K high and V high All are intermediate computational quantities; Mamba3(·), Mamba4(·), and Mamba5(·) represent three independent Mamba blocks.

7. The flight trajectory prediction method based on multi-channel frequency as described in claim 6, characterized in that, Step S4 also includes: Mamba0(·) and Mamba3(·) are the same Mamba block.

8. The flight trajectory prediction method based on multi-channel frequency as described in claim 6, characterized in that, Step S5 specifically includes: Low-frequency key information A low The low-frequency signal F is obtained by sequentially performing convolution and linear processing. low , F low =Linear(Conv l (A low )); In the formula, Conv l (·) indicates the third convolutional layer; For high-frequency key information A high The high-frequency signal F is obtained by sequentially performing convolution and linear processing. high , F high =Linear(Conv h (A high )); In the formula, Conv h (·) indicates the fourth convolutional layer.

9. The flight trajectory prediction method based on multi-channel frequency as described in claim 8, characterized in that, Step S6 specifically includes: Y 1:t+1 =IDWT(F low ,F high ) In the formula, Y 1:t+1 This represents a time-domain trajectory sequence, and IDWT(·) represents the inverse wavelet transform. Extract the last point Y t+1 Let P be the trajectory prediction point.

Images