Aircraft target trajectory prediction method and device based on multi-scale Mama

Abstract

The invention discloses an aircraft target trajectory prediction method and device based on multi-scale Mama, and belongs to the field of flight trajectory prediction, and the method comprises the steps: S1, carrying out the data preprocessing of trajectory sequence data, and recording the processed trajectory sequence as Xt: t + n; s2, converting Xt: t + n into a track sequence T which learns a change relation of time steps; s3, embedding the T code into a high-dimensional space, and capturing an interdependency relationship S of different components in a high-dimensional embedded vector; s4, compressing S at different scales, extracting features of different resolutions, and extracting to obtain a combined feature M of the two features; and S5, learning attention distribution of the joint feature M by using a wavelet attention module, and restoring a historical trajectory sequence Yt: t + n and a predicted trajectory point Yt + n + 1 through a linear layer and inverse wavelet transform. According to the method, high prediction precision and high response speed are balanced at the same time, and reliable technical support is provided for airspace management.

Description

Technical Field

[0001] The present invention relates to the field of flight trajectory prediction, and more specifically, to a method and device for predicting the target trajectory of an aircraft based on multi-scale Mamba. Background Art

[0002] With the continuous growth of air transportation volume and the increasing air traffic flow, how to accurately and quickly predict the future trajectory of an aircraft target has become an important topic in the aviation field. Trajectory prediction technology has wide applications in fields such as air traffic flow management, route optimization and fuel efficiency, unmanned aircraft and new air traffic, complex weather and special airspace, etc. Its accuracy directly affects the formulation of flight plans, the utilization efficiency of airspace, and the guarantee of flight safety.

[0003] Most of the existing aircraft target prediction technologies adopt a Transformer-like architecture to process flight trajectory data, taking advantage of its superiority in temporal feature extraction and complex sequence modeling. However, the self-attention mechanism of Transformer will generate quadratic computational complexity when dealing with long sequences, causing the computational amount to grow exponentially with the length of the input sequence. This complexity manifests as a high computational overhead and a long delay in aircraft target prediction that requires quick response, making it difficult to meet the real-time requirements, thus limiting the application of the Transformer architecture. Summary of the Invention

[0004] The purpose of the present invention is to overcome the deficiencies of the prior art and provide a method and device for predicting the target trajectory of an aircraft based on multi-scale Mamba, which balances high prediction accuracy and high response speed and provides reliable technical support for airspace management.

[0005] The purpose of the present invention is achieved through the following solutions:

[0006] A method for predicting the target trajectory of an aircraft based on multi-scale Mamba includes the following steps:

[0007] Step S1: Perform data preprocessing on the trajectory sequence data, and denote the processed trajectory sequence as X t:t+n ;

[0008] Step S2: Use intra-channel interaction to convert X t:t+n into a trajectory sequence T that has learned the variation relationship of time steps;

[0009] Step S3: Embed T into a high-dimensional space, and use an inter-channel interaction module to capture the mutual dependence relationship S among different components in the high-dimensional embedding vector;

[0010] Step S4: Use multi-scale extraction and fusion to compress S at different scales. First, extract features at different resolutions through the Mamba module, and then extract the combined feature M of the two;

[0011] Step S5: Use the wavelet attention module to learn the attention distribution of the combined feature M, and then restore the historical trajectory sequence Y through a linear layer and inverse wavelet transform t:t+n and the predicted trajectory point Y t+n+1 .

[0012] Furthermore, step S1 specifically includes sub-steps:

[0013] S11: Perform operations to remove missing values and duplicates on the trajectory dataset, and classify and filter according to conditions;

[0014] S12: Perform piecewise linear interpolation and resampling on the trajectory data to unify the time granularity;

[0015] S13: Slice and divide the trajectory through a sliding window operation. The slice length is n, and the slice step is 1. Cut trajectory sequences of different lengths into a set of trajectory slices of the same size, that is, the trajectory sequence {X t:t+n}

[0016] Furthermore, step S2 specifically includes sub-steps:

[0017] S21: Transpose X t:t+n , and sequentially pass through a linear layer and a ReLU activation layer to learn the change relationship between different time steps, and then transpose to unify the size of the vector:

[0018]

[0019] S22: Perform a residual connection on D that has learned the time step change information and the original input trajectory sequence X t:t+n so that the original trajectory can contain information of different time steps:

[0020] T = D + X t:t+n .

[0021] Furthermore, step S3 specifically includes:

[0022] S31: Use the Embedding method to embed T into a high-dimensional space to obtain a high-dimensional vector E;

[0023] E = Embedding(T);

[0024] S32: Successively pass the embedding vector E through average pooling, two 1D convolutions, and corresponding activation functions to learn the weights S' of each component of the embedding vector in the high-dimensional space; among them, the previous convolution reduces the dimension of the vector by half, and the subsequent convolution restores the vector to the original high-dimensional space:

[0025] s′ = ReLU(Conv1d 1 (avgPool(E)));

[0026] S′ = Sigmoid(Conv1d 2 (S′));

[0027] S33: Perform a dot product of E and the learned weights S′ to make the model pay attention to the importance ratio of different components in the high-dimensional vector:

[0028] S = S′·E.

[0029] Furthermore, step S4 specifically includes sub-steps:

[0030] S41: Pass s through two modules with roughly the same structure for compression at different scales; among them, 1D convolution with a kernel size of 1 is used for fine-grained compression, and 1D convolution with a kernel size of 3 is used for coarse-grained compression:

[0031] c 1 = ReLU(Conv1d 1 (avgPool(S)));

[0032] C 2 = ReLU(Conv1d 2 (avgPool(S)));

[0033] S42: Use the Mamba module to perform sequence modeling on the vectors compressed at different resolutions respectively, and perform residual connection:

[0034] M 1 = Mamba(C 1 ) + C 1 ;

[0035] M 2 = Mamba(C 2 ) + C 2 ;

[0036] S43: Concatenate the results of the two sequence modelings, that is, concatenate M 1 and M 2 to make the concatenated vector M′ have sequence modeling of both resolutions at the same time:

[0037] M′ = cat(C1 , C 2 );

[0038] S44: Model M' using the Mamba module, and simultaneously learn the hidden features at different resolutions, and then perform residual connection:

[0039] M = Mamba(M') + M'.

[0040] Furthermore, step S5 specifically includes sub-steps:

[0041] S51: Pass M through the wavelet attention WAtt module to obtain the context embedding C of the trajectory; wherein, in the enhancement stage, a weight matrix is calculated, representing the features facing different resolutions; in the convolution stage, the weight matrix is used to extract the context embedding for generating wavelet components from M and align it with the wavelet components to the same time dimension:

[0042] C = Conv1d(EnhancedBlock(M));

[0043] S52: Map C through the normalization layer and the linear layer to generate the wavelet component Q:

[0044] Q = FC(LN(C));

[0045] S53: Use the inverse wavelet transform module to restore the wavelet component Q to the real trajectory sequence Y t:t+n+1 ; where Y t:t+n is the reconstructed historical trajectory, and Y t+n+1 is the predicted next trajectory point:

[0046] Y t:t+n+1 = IDWT(Q).

[0047] Furthermore, in step S11, the operation of removing missing values and duplicates from the trajectory dataset and classifying and filtering according to conditions specifically includes sub-steps: Extracting the trajectory sequence data according to the flight number and removing the overly short trajectories with less than 20 points.

[0048] Furthermore, in step S12, the operation of performing piecewise linear interpolation and resampling on the trajectory data to unify the time granularity specifically includes sub-steps: Performing piecewise linear interpolation on the trajectory data and resampling at set time intervals to unify the time granularity.

[0049] An aircraft target trajectory prediction device based on multi-scale Mamba includes a processor and a memory, and a computer program is stored in the memory. When the computer program is loaded by the processor, it executes the method described in any one of the above.

[0050] The beneficial effects of the present invention include:

[0051] In the inventive concept, first, based on the Mamba architecture, the overhead required for computing is effectively reduced. Second, the Mamba architecture reduces redundant operations during the computing process and optimizes the use of model parameters by introducing a lightweight feature extraction module and an efficient temporal relationship modeling method. This technical solution not only reduces the overall computational complexity but also maintains high accuracy in flight trajectory prediction, thereby significantly improving the response speed of the system and adapting to airspace management and flight safety guarantee applications with high real-time requirements.

[0052] In the inventive concept, a new prediction model is proposed. This prediction model compresses trajectory vectors that have learned time step information and the proportion of component importance at different scales, and uses the Mamba module to perform sequence modeling on vectors at different scales, thereby mining hidden flight pattern features for subsequent trajectory reconstruction and prediction, improving the accuracy and speed of aircraft target trajectory prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the following-described drawings are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0054] Figure 1 It is a structural block diagram of an aircraft target trajectory prediction model based on multi-scale Mamba in an embodiment of the present invention;

[0055] Figure 2 It is a structural block diagram of the Mamba module in an embodiment of the present invention;

[0056] Figure 3 It is a structural block diagram of the wavelet attention WAtt module in an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0057] All the features disclosed in all the embodiments in this specification, or all the steps in the disclosed methods or processes, except for mutually exclusive features and / or steps, can be combined and / or extended and replaced in any way.

[0058] In a preferred embodiment, the present invention specifically proposes a method for predicting an aircraft target trajectory based on multi-scale Mamba. This method uses preprocessed trajectory data, uses the Mamba module to model trajectory vectors at different compression scales, and performs subsequent trajectory reconstruction and prediction. As Figure 1 shown, it specifically includes the following steps:

[0059] Step S1: Perform data preprocessing on the trajectory sequence data, and denote the processed trajectory as X t:t+n ;

[0060] Step S2: Use the intra-channel interaction module to convert X t:t+n into T that has learned the change relationship of time steps;

[0061] Step S3: Embed T into a high-dimensional space, and use the inter-channel interaction module to capture the mutual dependence relationship S of different components in the high-dimensional embedding vector;

[0062] Step S4: Use the multi-scale extraction and fusion module to compress S at different scales. First, extract features of different resolutions through Mamba, and then extract the joint feature M of the two;

[0063] Step S5: Use the wavelet attention module to learn the attention distribution of the joint feature M, and then restore the historical trajectory sequence Y t:t+n and the predicted trajectory point Y t+n+1 .

[0064] In a further embodiment, step S1 specifically includes sub-steps:

[0065] S11: Perform missing value removal and duplicate removal operations on the trajectory data set, and classify and filter according to conditions; in this embodiment, extract the trajectory sequence data according to the flight number, and remove the overly short trajectories with less than 20 points;

[0066] S12: Perform piecewise linear interpolation and resampling on the trajectory data to unify the time granularity; in this embodiment, perform piecewise linear interpolation and resampling operations at intervals of 10 s on the trajectory data to unify the time granularity;

[0067] S13: Slice and divide the trajectory through a sliding window operation to obtain a trajectory slice set {X t:t+n}; in this embodiment, the slice length is n, the slice step size is 1, and the trajectory sequences of different lengths are sliced into the same-size trajectory slice set {X t:t+n}.

[0068] In a further embodiment, step S2 specifically includes sub-steps:

[0069] S21: Transpose the trajectory data X t:t+n , and successively pass through a linear layer and a ReLU activation layer to learn the change relationship between different time steps, and then transpose to unify the size of the vector:

[0070]

[0071] S22: Residually connect D that has learned the time step change information with the original input trajectory sequence X t:t+n so that the original trajectory can contain information from different time steps:

[0072] T = D + X t:t+n ;

[0073] In a further embodiment, step S3 specifically includes:

[0074] S31: Use the Embedding method to embed T into a high-dimensional space to obtain a high-dimensional vector E; in this embodiment, the trajectory sequence T containing time information is embedded into a high-dimensional space using the Embedding method to obtain the expansion of the low-dimensional attributes in the high-dimensional space:

[0075] E = Embedding(T);

[0076] S32: Pass the embedded vector E through average pooling, two 1D convolutions, and corresponding activation functions in sequence to learn the weights S' of each component of the embedded vector in the high-dimensional space. Among them, the previous convolution reduces the dimension of the vector by half, and the latter convolution restores the vector to the original high-dimensional space:

[0077] S' = ReLU(Conv1d 1 (avgPool(E)));

[0078] S' = Sigmoid(Conv1d 2 (S'));

[0079] S33: Multiply E by the previously learned weights S' to make the model pay attention to the importance ratio of different components in the high-dimensional vector:

[0080] S = S' · E;

[0081] In a further embodiment, step S4 specifically includes sub-steps:

[0082] S41: Pass S through two modules with roughly the same structure for compression at different scales. Among them, the 1D convolution with a kernel size of 1 is used for fine-grained compression, and the 1D convolution with a kernel size of 3 is used for coarse-grained compression:

[0083] C 1 = ReLU(Conv1d 1 (avgPool(S)));

[0084] C 2 = ReLU(Conv1d 2 (avgPool(S)));

[0085] S42: Use the Mamba module to perform sequence modeling on vectors compressed at different resolutions respectively, and perform residual connection. The structural diagram of Mamba is as shown in Figure 2 the figure. It provides advanced modeling capabilities similar to Transformers, but does not introduce quadratic computational complexity and can have equally excellent performance when dealing with long time-series data:

[0086] M 1 = Mamba(C 1 ) + C 1

[0087] M 2 = Mamba(C 2 ) + C 2

[0088] S43: Concatenate the results of the two sequence modellings, that is, concatenate M 1 and M 2 so that the concatenated vector M' has sequence modellings of two resolutions at the same time:

[0089] M' = cat(C 1 , C 2 )

[0090] S44: Use the Mamba module to model M', learn hidden features at different resolutions at the same time, and then perform residual connection:

[0091] M = Mamba(M') + M';

[0092] In a further embodiment scheme, step S5 specifically includes sub-steps:

[0093] S51: Pass the trajectory sequence modeling M through the wavelet attention WAtt module to obtain the context embedding C of the trajectory. The structural diagram of the wavelet attention WAtt module is as shown in Figure 3 the figure, which includes two stages: the enhancement stage and the convolution stage. The enhancement stage calculates a weight matrix representing features facing different resolutions; the convolution stage uses the weight matrix to extract the context embedding for generating wavelet components from M and aligns it with the wavelet components to the same time dimension:

[0094] C = Conv1d(EnbancedBlock(M))

[0095] S52: Map C through a normalization layer and a linear layer to generate the wavelet component Q:

[0096] Q = FC(LN(C))

[0097] S53: Use the inverse wavelet transform module to restore the wavelet component Q to the true trajectory sequence Y t:t+n+1 . Where Y t:t+n is the reconstructed historical trajectory, and Y t+n+1 is the predicted next trajectory point:

[0098] Y t:t+n+1 = IDWT(Q).

[0099] The units involved in the embodiments of the present invention can be implemented in software or in hardware, and the described units can also be provided in a processor. Among them, the names of these units do not constitute a limitation to the unit itself in some cases.

[0100] According to one aspect of the embodiments of the present invention, there is provided a computer program product or a computer program, which includes computer instructions stored in a computer-readable storage medium. The processor of the computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device executes the methods provided in the above various optional implementation manners.

[0101] As another aspect, the embodiments of the present invention further provide a computer-readable medium, which may be included in the electronic device described in the above embodiments; or may exist alone without being assembled into the electronic device. The above computer-readable medium carries one or more programs, and when the one or more programs are executed by an electronic device, the electronic device implements the methods described in the above embodiments.

Claims

1. A method for predicting aircraft target trajectories based on multi-scale Mamba, characterized in that: The following steps are involved: Step S1: Preprocess the trajectory sequence data and record the processed trajectory sequence as X t:t+n ; Step S2: Use intra-channel interaction to convert X t:t+n Convert to a trajectory sequence T that learns the changing relationship of time steps; Step S3: embed the T code into a high-dimensional space and use the channel-to-channel interaction module to capture the interdependencies S of different components in the high-dimensional embedding vector; Step S4: Use multi-scale extraction and fusion to compress S at different scales, extract features of different resolutions through the Mamba module, and then extract the joint feature M of the two; Step S5: Use the wavelet attention module to learn the attention distribution of the joint feature M, and then restore the historical trajectory sequence Y through the linear layer and inverse wavelet transform t:t+n and predicted trajectory point Y t+n+1 .

2. The method for predicting aircraft target trajectory based on multi-scale Mamba according to claim 1, characterized in that: Step S1 specifically includes sub-steps: S11: remove missing and duplicate data from the trajectory data set, and classify and filter it according to conditions; S12: Perform piecewise linear interpolation and resampling on the trajectory data to unify the time granularity; S13: Slice the trajectory through a sliding window operation, with a slice length of n and a slice step of 1, and slice the trajectory sequences of different lengths into a set of trajectory slices of the same size, i.e., the trajectory sequence {X t:t+n }.

3. The method for predicting aircraft target trajectory based on multi-scale Mamba according to claim 1, characterized in that: Step S2 specifically includes the following sub-steps: S21: x t:t+n Transpose, pass through the linear layer and ReLU activation layer in turn, learn the change relationship between different time steps, and then transpose to unify the size of the vector: S22: D, which has learned the time step change information, is combined with the original input trajectory sequence X t:t+n Perform residual connection so that the original trajectory can contain information at different time steps: T=D+X t:t+n 。 4. The method for predicting aircraft target trajectory based on multi-scale Mamba according to claim 1, characterized in that: Step S3 specifically includes: S31: Use the Embedding method to embed T into a high-dimensional space to obtain a high-dimensional vector E; E = Embedding(T); S32: The embedded vector E is subjected to average pooling, two one-dimensional convolutions, and corresponding activation functions in turn to learn the weights S′ of each component of the embedded vector in the high-dimensional space; the first convolution reduces the dimension of the vector by half, and the second convolution restores the vector to the original high-dimensional space: S′=ReLU(Conv1d1(avgPool(E))); S′=Sigmoid(Conv1d2(S′)); S33: Multiply E by the learned weight S′ to make the model pay attention to the importance of different components in the high-dimensional vector: S=S′·E.

5. The method for predicting aircraft target trajectory based on multi-scale Mamba according to claim 1, characterized in that: Step S4 specifically includes the following sub-steps: S41: S is compressed at different scales through two modules with roughly the same structure. Among them, the one-dimensional convolution with a convolution kernel size of 1 is used for fine-grained compression, and the one-dimensional convolution with a convolution kernel size of 3 is used for coarse-grained compression: C1=ReLU(Conv1d1(avgPool(S))); C2=ReLU(Conv1d2(avgPool(S))); S42: Use the Mamba module to model the sequences of vectors compressed at different resolutions and perform residual connections: M1=Mamba(C1)+C1; M2=Mamba(C2)+C2; S43: Concatenate the results of the two sequence modeling, that is, concatenate M1 and M2, so that the concatenated vector M′ has sequence modeling of two resolutions at the same time: M′=cat(C1,C2); S44: Use the Mamba module to model M′ and learn hidden features at different resolutions, followed by residual connections: M = Mamba (M') + M'.

6. The method for predicting aircraft target trajectory based on multi-scale Mamba according to claim 1, characterized in that: Step S5 specifically includes the following sub-steps: S51: M is passed through the wavelet attention WAtt module to obtain the context embedding C of the trajectory; wherein, a weight matrix is ​​calculated in the enhancement stage to represent the features facing different resolutions; in the convolution stage, the weight matrix is ​​used to extract M into the context embedding used to generate the wavelet component, and aligned with the wavelet component to the same time dimension: C=Conv1d(EnhancedBlock(M)); S52: Map C through the normalization layer and the linear layer to generate the wavelet component Q: Q = FC(LN(C)); S53: Use the inverse wavelet transform module to restore the wavelet component Q to the real trajectory sequence Y t:t+n+1 ; where Y t:t+n That is the historical trajectory of reconstruction, Y t+n+1 This is the predicted next trajectory point: Y t:t+n+1 =IDWT(Q)。 7. The method for predicting aircraft target trajectory based on multi-scale Mamba according to claim 2, characterized in that: In step S11, the trajectory data set is subjected to deletion and duplication removal operations, and is classified and screened according to conditions, specifically including the sub-steps of extracting trajectory sequence data according to flight numbers, and removing short trajectories with less than 20 points.

8. The method for predicting aircraft target trajectory based on multi-scale Mamba according to claim 2, characterized in that: In step S12, the piecewise linear interpolation and resampling of the trajectory data to unify the time granularity specifically includes sub-steps: piecewise linear interpolation of the trajectory data and resampling operations at intervals of a set time length to unify the time granularity.

9. An aircraft target trajectory prediction device based on multi-scale Mamba, characterized in that: The method comprises a processor and a memory, wherein a computer program is stored in the memory, and when the computer program is loaded by the processor, the method according to any one of claims 1 to 8 is executed.

Images