Aircraft trajectory prediction method in multi-dimensional variable scene based on multi-attention collaborative network

Abstract

The invention discloses an aircraft trajectory prediction method in a multi-dimensional variable scene based on a multi-attention collaborative network, and belongs to the technical field of trajectory prediction. According to the method, the problem of poor prediction accuracy caused by difficulty in fully modeling a complex relationship between variables in an existing prediction method based on data driving in a multivariable scene is solved. The variable distillation attention network designed by the invention weakens the negative influence of non-predictive variables, and converts the current sequence into spatial semantics. A new KeLSTM network variant is also designed to calculate the interaction between variables, alleviate information conflicts, and distinguish the time influence of the variables on tasks. According to the method, the negative influence of irrelevant variables can be reduced, the contribution of important variables to prediction is emphasized, and potential interaction among past target variables, non-predictive variables and an upcoming target, especially subtle differences between the influences of the target and the non-predictive variables, is successfully captured. The method can be applied to aircraft trajectory prediction.

Description

Technical Field

[0001] The present invention belongs to the technical field of trajectory prediction, and in particular relates to a method for predicting aircraft trajectories in a multi-dimensional variable scenario based on a multi-attention collaborative network. Background Art

[0002] Simulation entities are the basic functional units that constitute distributed simulation systems. The main purpose of distributed simulation is to build a real and perceptually consistent virtual environment among the participating simulation entities. When the state of an entity in the experiment changes or its position in the virtual environment changes, all participating entities can promptly discover the update of the entity state. However, due to the limitation of network communication capacity and the uncertain delay of information transmission, different simulation entities often fail to reach a consistent observation of the same entity state update, which leads to the problem of time and space inconsistency in the distributed simulation process. As distributed simulation systems develop towards large scale, many entities, wide regions, different structures, and long time, the simulation process is becoming more and more detailed, and the problem of time and space inconsistency in distributed simulation systems is becoming more and more prominent. Large-scale distributed simulation systems have very high requirements for time and space consistency. The existence of time and space inconsistency will lead to inconsistent cognition among simulation entities, resulting in adverse effects such as causal reversal and misunderstanding, which will destroy the realism and effectiveness of the simulation and affect the credibility of the simulation results.

[0003] In order to solve the problem of time and space inconsistency, the dead reckoning positioning algorithm uses prediction technology between entities to achieve compensation for time and space inconsistency caused by unstable network transmission, thereby ensuring the authenticity of the simulation and the credibility of the results. However, it only relies on short-term historical data to achieve entity state prediction, and it is difficult to mine the motion characteristics of various high-maneuverability targets from the time and space dimensions. It cannot handle the sudden changes in the motion behavior of the predicted object, which affects the accuracy of the entity state prediction. Therefore, other more efficient and accurate trajectory prediction methods are needed to predict the trajectory of simulated entities such as aircraft in distributed simulation systems.

[0004] Most of the existing aircraft trajectory prediction methods are based on predefined motion models and use algorithms such as Markov chains and Kalman filters for trajectory prediction. However, when performing tasks, aircraft are often in complex nonlinear environments. Their motion behavior is not only constrained by the aircraft's own dynamics, but also affected by a variety of external factors such as meteorological conditions, topography, and environmental interference. These factors make the motion state of the aircraft difficult to predict. When facing such a complex environment, traditional physical models are difficult to accurately describe the actual motion behavior of the aircraft, resulting in a decrease in prediction accuracy. In addition, the construction of physical models usually relies on a large number of assumptions and simplifications. When these assumptions are no longer valid, the applicability and accuracy of the model will also decrease. In recent years, with the development of machine learning and deep learning methods, it has become feasible to use data-driven methods to learn the complex motion laws of aircraft to improve the accuracy of trajectory prediction. Therefore, trajectory tracking and prediction algorithms based on deep neural networks have attracted more and more attention from scholars.

[0005] Data-driven trajectory prediction methods rely on a large amount of historical data and real-time data to predict the motion of aircraft, and have significant advantages in dealing with complex nonlinear environments. Data-driven methods learn from a large amount of historical data and do not require a deep understanding of the physical characteristics of the aircraft. This not only reduces the difficulty of modeling, but also improves the adaptability of the model, and can effectively cope with changing flight mission scenarios. However, data-driven methods still have significant shortcomings in multivariate prediction. When dealing with multivariate data, these methods are difficult to fully model the complex relationships between variables, and often require complex feature selection, otherwise it is easy to cause information loss or noise introduction, resulting in poor prediction accuracy.

[0006] Therefore, developing an aircraft trajectory prediction technology in a multi-dimensional variable scenario is of great significance for ensuring anti-interference performance, improving the spatiotemporal consistency of the aircraft state recognized by various resources during virtual-reality collaborative testing, and reducing the impact of aircraft state recognition lag on the realism of complex environment construction. Summary of the invention

[0007] The purpose of the present invention is to solve the problem that the existing data-driven prediction methods in multi-variable scenarios are difficult to fully model the complex relationship between variables, resulting in poor prediction accuracy, and propose a method for predicting aircraft trajectories in multi-dimensional variable scenarios based on a multi-attention collaborative network.

[0008] The technical solution adopted by the present invention to solve the above technical problems is: a method for predicting aircraft trajectories in a multi-dimensional variable scenario based on a multi-attention collaborative network, the method specifically comprising the following steps:

[0009] Step 1: Collect M sets of trajectory data of the aircraft, and each set of trajectory data includes time data, speed data, position data and heading angle data;

[0010] After preprocessing each set of collected trajectory data respectively, the position data in each set of preprocessed trajectory data is reconstructed to obtain each set of processed trajectory data;

[0011] Step 2: cluster each group of processed trajectory data to obtain each type of trajectory data;

[0012] Step 3: Build a prediction model based on multi-attention collaborative network for each type of trajectory data. The prediction model includes two branches, T-net and NP-net, and use each type of trajectory data as the training set of the corresponding prediction model.

[0013] After training each prediction model using the training set, each trained prediction model is obtained;

[0014] Step 4: Use the data to be predicted as the input of the corresponding trained prediction model to obtain the trajectory prediction result.

[0015] Furthermore, the collected trajectory data are preprocessed separately, specifically:

[0016] Invalid values ​​in the speed data, position data and heading angle data are removed, and noise in the speed data, position data and heading angle data is filtered.

[0017] Furthermore, the position data in each set of preprocessed trajectory data is reconstructed, specifically:

[0018] Take the position data in the m'th group of trajectory data as an example:

[0019] Step 11: Set the distance threshold;

[0020] Step 12: Calculate the distance between the position data corresponding to two adjacent time points;

[0021] If the distance between the position data corresponding to two adjacent time points is greater than the set distance threshold, the two adjacent time points are marked as discontinuity points;

[0022] If the distance between the position data corresponding to two adjacent time points is less than or equal to the set distance threshold, no processing is required;

[0023] After respectively judging the position data corresponding to every two adjacent time points in the m'th group of trajectory data, all discontinuity points in the position data of the m'th group of trajectory are obtained;

[0024] Step 13, respectively obtain the number of data points between every two adjacent discontinuity points, and compare the obtained number of data points with the trajectory noise threshold;

[0025] If the number of data points between two adjacent discontinuity points is greater than the trajectory noise threshold, the trajectory data corresponding to the two adjacent discontinuity points and the data points between the two adjacent discontinuity points are saved as a new set of trajectory data;

[0026] If the number of data points between two adjacent discontinuity points is less than or equal to the trajectory noise threshold, the trajectory data corresponding to the two adjacent discontinuity points and the data points between the two adjacent discontinuity points are regarded as noise removal;

[0027] Step 14: remove the new trajectory data saved in step 13 and the noise data removed in step 13 from the m'th group of trajectory data, and use the remaining data as the updated m'th group of trajectory data;

[0028] After the position data in each set of trajectory data are processed respectively, the updated trajectory data and the saved new trajectory data are used as each set of processed trajectory data.

[0029] Furthermore, the clustering of each group of processed trajectory data is performed based on the position data in each group of trajectory data, and the clustering process is:

[0030] Step 21: Calculate the distance matrix D based on the position data in each set of processed trajectory data:

[0031]

[0032] Among them, the trajectory set E t is a set of processed trajectory data. tj′ is the set of trajectories E t The j′th group of trajectory data in E, j′=1,2,…,N, N is the set E t The number of groups of trajectory data in It is the set E t The similarity between the first group of trajectory data and the j′th group of trajectory data;

[0033] Step 22: Initialize the number of cluster centers i′=1;

[0034] Step 2: According to the distance matrix D, the trajectory set E t Each group of trajectory data in is clustered into i′ class;

[0035] Step 24: Calculate the clustering error SSE according to the clustering result. If the clustering error converges, the final clustering result is obtained. If the clustering error does not converge, set i′=i′+1 and return to execute step 23.

[0036] Furthermore, the set E t The calculation method of the similarity between the first group of trajectory data and the j′th group of trajectory data is:

[0037]

[0038] Among them, h(E t1 ,E tj′ ) is an intermediate variable, a is the position point in the first set of trajectory data, and b is the position point in the j′th set of trajectory data;

[0039]

[0040] Furthermore, the calculation process of the clustering error SSE is:

[0041]

[0042] Among them, C i′ represents the i′th cluster, and p represents C i′ The position data in any set of trajectory data, m i′ For C i′ The centroid of the position data in all group trajectory data, k′ is the total number of categories.

[0043] Furthermore, the training process of the prediction model is:

[0044] Take any prediction model as an example

[0045] Step 31: Use each set of trajectory data in the training set of the current prediction model to construct a set of target variable sequences and non-prediction variable sequences;

[0046] Among them, a set of target variable sequences and non-predictor variable sequences are represented as and x t is the non-predictor variable corresponding to the t-th time point in the trajectory data, They are x t The first, second, and nth variables in ;

[0047] Where T is the number of time points in a set of trajectory data;

[0048] Step 3.2: Sequence non-predictor variables As the input of the NP-net branch, the hidden state at each time point is output through the NP-net branch

[0049] Step 33: The T-net branch includes an encoder module and a decoder module, the encoder module includes T knowledge-enhanced LSTM units, and the decoder module includes Δ knowledge-enhanced LSTM units;

[0050] Output the hidden state of the NP-net branch and the target variable sequence As the input of the encoder module, the target variable sequence is output through the encoder module The hidden state sequence of

[0051] Will hide the sequence As the input of the decoder module, the decoder module outputs the hidden state based on the temporal attention mechanism, passes the hidden state output by the decoder module through the multi-layer perceptron, and outputs the prediction result through the multi-layer perceptron;

[0052] Step 34: Calculate the loss function according to the prediction result output in step 33, adjust the parameters of the prediction model based on the loss function calculation result and the back propagation learning algorithm, and stop training when the set maximum number of training times is reached to obtain a trained prediction model.

[0053] Furthermore, the NP-net branch includes a variable distillation attention module and an LSTM module, wherein the variable distillation attention module includes T variable distillation attention units, and the LSTM module includes T LSTM units;

[0054] The working process of the NP-net branch is as follows:

[0055] Each variable distillation attention unit includes two layers of forward neural networks. The first layer of the forward neural network of the t-th variable distillation attention unit converts x t Each variable in is mapped to the m-dimensional space, and x t The hidden representation of each variable in :

[0056]

[0057] Among them, tanh is the activation function, yes The hidden representation of k = 1, 2, ..., n, is the embedding vector, is the bias parameter;

[0058] The second layer of the feed-forward neural network is for each hidden representation Assign weights:

[0059]

[0060] in, It is hidden representation The weight of is a variable-order context vector;

[0061] based on and x t Convert to spatial semantic sequence

[0062]

[0063] Spatial semantic sequence As the input of the t-th LSTM unit, the hidden state is output by the t-th LSTM unit

[0064]

[0065] in, is the hidden state output by the t-1th LSTM unit, is the hidden state output by the t-th LSTM unit, i t is the output of the input gate of the t-th LSTM unit, f t is the output of the forget gate of the t-th LSTM unit, o t is the output of the output gate of the t-th LSTM unit, σ represents the sigmoid activation function, and is the bias parameter, ⊙ is the element-wise multiplication, W i , W f , W o and W c is the cycle weight, U i , U f , U o and U c is the input weight, c t is the state of the t-th LSTM unit, c t-1 is the state of the t-1th LSTM unit, is an intermediate variable.

[0066] Furthermore, the working process of the T-net branch is as follows:

[0067] Step 1: Hide the state and the target variable sequence y t As the input of the tth knowledge-enhanced LSTM unit in the encoder module, that is, but

[0068]

[0069] in, is the hidden state output by the t-1th knowledge-enhanced LSTM unit in the encoder module, is the candidate state of the tth knowledge-enhanced LSTM unit in the encoder module;

[0070] According to the hidden state and Generate the hidden state of the tth knowledge-enhanced LSTM unit:

[0071]

[0072] in, represents the transpose of a matrix, is the weight parameter, is the auxiliary context vector, s t yes The importance score of t yes The weight of

[0073]

[0074] Among them, W a and W h is the weight parameter, is the hidden state output by the tth knowledge-enhanced LSTM unit in the encoder module;

[0075] Step 2: As the input of each knowledge-enhanced LSTM unit in the decoder module, for the τth knowledge-enhanced LSTM unit in the decoder module, based on the temporal attention mechanism, and the hidden state d of the τ-1th knowledge-enhanced LSTM unit in the decoder module τ-1 Generate context vector hc τ :

[0076]

[0077] in, and is the weight parameter, ⊙ is the element-wise multiplication, e t yes The importance score, β t yes The attention weight of

[0078] According to β t right Perform weighted summation to generate context vector hc τ :

[0079]

[0080] Step 3: Pass the hidden state of the τ-1th knowledge-enhanced LSTM unit through the multilayer perceptron, output the prediction result of the τ-1th knowledge-enhanced LSTM unit through the multilayer perceptron, and compare the prediction result of the τ-1th knowledge-enhanced LSTM unit with the context vector hc τ Perform cascade operations;

[0081] The cascade operation result and hidden state d τ-1 and hidden state As the input of the τth knowledge-enhanced LSTM unit, that is,

[0082] Among them, [:] is a cascade operation, is the prediction result of the τ-1th knowledge-enhanced LSTM unit in the decoder module, d τ is the hidden state of the τth knowledge-enhanced LSTM unit in the decoder module;

[0083] The hidden state d τ After passing through a multi-layer perceptron, the prediction result of the τth knowledge-enhanced LSTM unit is output through the multi-layer perceptron.

[0084] Furthermore, the hidden state d τ After passing through the multi-layer perceptron, the prediction result of the τth knowledge-enhanced LSTM unit is output by the multi-layer perceptron, specifically:

[0085]

[0086] in, is the weight parameter, is the prediction result of the τth knowledge-enhanced LSTM unit.

[0087] The beneficial effects of the present invention are:

[0088] The present invention first designs a variable distillation attention mechanism as the first layer network. The variable distillation attention network weakens the negative impact of non-predicted variables and converts the current sequence into spatial semantics. In addition, a new long short-term memory network variant KeLSTM is designed to calculate the interaction between variables, alleviate information conflicts, and distinguish the temporal impact of variables on tasks. The method of the present invention can reduce the negative impact of irrelevant variables, emphasize the contribution of important variables to prediction, and successfully capture the potential interactions between past target variables, non-predicted variables and upcoming targets, especially the subtle differences between the impacts of targets and non-predicted variables. The method of the present invention can improve the accuracy of prediction without the need for complex feature selection, and the method of the present invention still has good adaptability in complex nonlinear environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0089] Figure 1 It is the workflow diagram of the prediction model;

[0090] In the present invention, y t As the χ t ,Will As the figure

[0091] Figure 2 It is a flow chart of a method for predicting aircraft trajectories in a multi-dimensional variable scenario based on a multi-attention collaborative network of the present invention. DETAILED DESCRIPTION

[0092] Specific implementation method 1: Combination Figure 2 The present embodiment describes a method for predicting aircraft trajectories in a multi-dimensional variable scenario based on a multi-attention collaborative network, and the method specifically includes the following steps:

[0093] Step 1: Collect M sets of trajectory data of the aircraft, and each set of trajectory data includes time data, speed data, position data and heading angle data;

[0094] After preprocessing each set of collected trajectory data respectively, the position data in each set of preprocessed trajectory data is reconstructed to obtain each set of processed trajectory data;

[0095] Step 2: cluster each group of processed trajectory data (clustering methods include but are not limited to k-medoids algorithm) to obtain each type of trajectory data; wherein each type of trajectory data corresponds to a trajectory route, the trajectory change trends in each class are similar, and the trajectory change trends between different classes are different.

[0096] Step 3: Build a prediction model based on multi-attention collaborative network (MACN network) for each type of trajectory data. The prediction model includes two branches, T-net and NP-net, and use each type of trajectory data as the training set of the corresponding prediction model.

[0097] After training each prediction model using the training set, each trained prediction model is obtained;

[0098] Step 4: Use the data to be predicted as the input of the corresponding trained prediction model (the trajectory route category of the data to be predicted is known, so the data to be predicted can be used as the input of the prediction model corresponding to the known trajectory route category) to obtain the trajectory prediction result.

[0099] The NP-net branch of the present invention mainly focuses on non-predicted variables, captures the most relevant variables and time dependencies; the T-net branch predicts the target variable, connects the encoder module and the decoder module to the NP-net branch, and thus utilizes the output of the NP-net branch to assist learning and decision-making; a knowledge-enhanced LSTM (KeLSTM) is used as the encoder module and the decoder module of the T-net branch to selectively enhance the potential representation of the target variable, handle the conflict between the target variable and the non-predicted variables, and improve the accuracy of the prediction.

[0100] Specific implementation method 2: This implementation method is different from the specific implementation method 1 in that each group of collected trajectory data is preprocessed separately, specifically:

[0101] Invalid values ​​in the speed data, position data and heading angle data are removed, and noise in the speed data, position data and heading angle data is filtered.

[0102] The other steps and parameters are the same as those in the first embodiment.

[0103] Specific implementation method three: This implementation method is different from specific implementation method one or two in that the position data in each group of preprocessed trajectory data is reconstructed, specifically:

[0104] Take the position data in the m'th group of trajectory data as an example:

[0105] Step 11: Set the distance threshold;

[0106] Step 12: Calculate the distance between the position data corresponding to two adjacent time points;

[0107] If the distance between the position data corresponding to two adjacent time points is greater than the set distance threshold, the two adjacent time points are marked as discontinuity points;

[0108] If the distance between the position data corresponding to two adjacent time points is less than or equal to the set distance threshold, no processing is required;

[0109] After respectively judging the position data corresponding to every two adjacent time points in the m'th group of trajectory data, all discontinuity points in the position data of the m'th group of trajectory are obtained;

[0110] Step 13, respectively obtain the number of data points between every two adjacent discontinuity points, and compare the obtained data points with the trajectory noise threshold;

[0111] If the number of data points between two adjacent discontinuity points is greater than the trajectory noise threshold, the trajectory data corresponding to the two adjacent discontinuity points and the data points between the two adjacent discontinuity points are saved as a new set of trajectory data;

[0112] If the number of data points between two adjacent discontinuity points is less than or equal to the trajectory noise threshold, the trajectory data corresponding to the two adjacent discontinuity points and the data points between the two adjacent discontinuity points are regarded as noise removal;

[0113] Step 14: remove the new trajectory data saved in step 13 and the noise data removed in step 13 from the m'th group of trajectory data, and use the remaining data as the updated m'th group of trajectory data;

[0114] After the position data in each set of trajectory data are processed respectively, the updated trajectory data and the saved new trajectory data are used as each set of processed trajectory data.

[0115] The other steps and parameters are the same as those in the first or second embodiment.

[0116] The distance threshold and trajectory noise threshold can be selected according to actual conditions.

[0117] Specific implementation method 4: This implementation method is different from the specific implementation methods 1 to 3 in that the clustering of each group of trajectory data after processing is performed based on the position data in each group of trajectory data, and the clustering process is:

[0118] Step 21: Calculate the distance matrix D based on the position data in each set of processed trajectory data:

[0119]

[0120] Among them, the trajectory set E t is a set of processed trajectory data. tj′ is the set of trajectories E t The j′th group of trajectory data in E, j′=1,2,…,N, N is the set E t The number of groups of trajectory data in It is the set E t The similarity between the first group of trajectory data and the j′th group of trajectory data;

[0121] Step 22: Initialize the number of cluster centers i′=1;

[0122] Step 2: According to the distance matrix D, the trajectory set E t Each group of trajectory data in is clustered into i′ class;

[0123] Step 24: Calculate the clustering error SSE based on the clustering results. If the clustering error converges (when i′=2, the convergence judgment is started, if i′=1, it is directly processed as non-convergence), the final clustering result is obtained. If the clustering error does not converge, set i′=i′+1 and return to step 23.

[0124] The other steps and parameters are the same as those in Specific Embodiments 1 to 3.

[0125] Clustering results are usually evaluated using the silhouette coefficient, which measures the quality of clustering from the perspective of intra-cluster compactness and inter-cluster separation, and derives an evaluation standard for clustering quality. The silhouette coefficient can be expressed as follows:

[0126]

[0127] Among them, a(x) is the average distance between x and other trajectories in the cluster to which x belongs. The value of a(x) reflects the compactness between the clusters to which x belongs. The smaller the value, the more compact the cluster. Calculate the average distance from x to all trajectories in another cluster. Similarly, calculate the average distance corresponding to each other cluster separately, and take the smallest average distance as b(x). The larger the value, the better the separation. The closer the silhouette coefficient is to 1, the better the clustering effect.

[0128] Specific implementation mode 5: This implementation mode is different from the specific implementation modes 1 to 4 in that the set E t The calculation method of the similarity between the first group of trajectory data and the j′th group of trajectory data is:

[0129]

[0130] Among them, h(E t1 ,E tj′ ) is an intermediate variable, a is the position point in the first group of trajectory data, and b is the position point in the j′th group of trajectory data; that is, first calculate the distance between the position point a in the first group of trajectory data and each position point b in the j′th group of trajectory data, and then select the minimum distance corresponding to the position point a. Similarly, calculate the minimum distance corresponding to each position point in the first group of trajectory data respectively, and the maximum value of all the minimum distances obtained is h(E t1 ,E tj′ );

[0131]

[0132] Calculation process and h(E tj′ ,E t1 )similar;

[0133]

[0134] The other steps and parameters are the same as those in Specific Embodiments 1 to 4.

[0135] Specific implementation method 6: This implementation method is different from the specific implementation methods 1 to 5 in that the calculation process of the clustering error SSE is:

[0136]

[0137] Among them, C i′ represents the i′th cluster, and p represents C i′ The position data in any set of trajectory data, m i′ For C i′ The centroid of the position data in all group trajectory data, k′ is the total number of categories.

[0138] The other steps and parameters are the same as those in Specific Implementations 1 to 5.

[0139] Specific implementation method seven: Combination Figure 1 The present embodiment is different from any one of the specific embodiments 1 to 6 in that the training process of the prediction model is as follows:

[0140] Take any prediction model as an example

[0141] Step 31: Use each set of trajectory data in the training set of the current prediction model to construct a set of target variable sequences and non-prediction variable sequences;

[0142] Among them, a set of target variable sequences and non-predictor variable sequences are represented as and x t is the non-predictor variable corresponding to the t-th time point in the trajectory data, They are x t The first, second, and nth variables in ;

[0143] Where T is the number of time points in a set of trajectory data;

[0144] For example, when the position data needs to be predicted, the target variable sequence is a sequence composed of the position data in the trajectory data, and the non-prediction variable sequence is a sequence composed of other data in the trajectory data except the position. That is to say, only one type of data can be predicted each time. The target variable sequence only includes the data that needs to be predicted, and the non-prediction variable sequence includes all the data that does not need to be predicted. During the training process, the target variable sequence needs to be constructed with each type of data and used for model training.

[0145] Step 3.2: Sequence non-predictor variables As the input of the NP-net branch, the hidden state at each time point is output through the NP-net branch

[0146] Step 33: The T-net branch includes an encoder module and a decoder module, the encoder module includes T knowledge-enhanced LSTM units (KeLSTM), and the decoder module includes Δ knowledge-enhanced LSTM units (KeLSTM);

[0147] Output the hidden state of the NP-net branch and the target variable sequence As the input of the encoder module, the target variable sequence is output through the encoder module The hidden state sequence of

[0148] Will hide the sequence As the input of the decoder module, the decoder module outputs the hidden state based on the temporal attention mechanism, passes the hidden state output by the decoder module through the multi-layer perceptron, and outputs the prediction result through the multi-layer perceptron;

[0149] Step 34: Calculate the loss function according to the prediction result output in step 33, adjust the parameters of the prediction model based on the loss function calculation result and the back propagation learning algorithm, and stop training when the set maximum number of training times is reached to obtain a trained prediction model.

[0150] The other steps and parameters are the same as those in Specific Embodiments 1 to 6.

[0151] KeLSTM first calculates the non-prediction sequence and the target variable y t Then, by selectively fusing the interactive information using a gated fusion unit, y is enhanced. t Potential manifestations of KeLSTM uses interactive attention to capture important latent representations It is used for prediction and balances the conflict between target variables and non-predicted variables through gated fusion units. Finally, the prediction model uses a multi-layer perceptron as the task learning layer (TLL). The task learning layer transforms the state d τ Convert to predicted value

[0152] The prediction model of the present invention uses the root mean square error (RMSE) as the loss function and uses the Adam optimizer to optimize the network parameters. The learning rate is initialized to 0.001 and reduced according to the training loss. The calculation formula of RMSE can be expressed as follows:

[0153]

[0154] Among them, Δ is the size of the prediction range, N is the number of samples, is the true value, is the corresponding predicted value.

[0155] And test the trained prediction model.

[0156] Specific implementation eight: This implementation is different from any one of specific implementations one to seven in that the NP-net branch includes a variable distillation attention module (VDM) and an LSTM module, wherein the variable distillation attention module includes T variable distillation attention units, and the LSTM module includes T LSTM units;

[0157] The working process of the NP-net branch is as follows:

[0158] Each variable distillation attention unit includes two layers of forward neural networks. The first layer of the forward neural network of the t-th variable distillation attention unit converts x t Each variable in is mapped to the m-dimensional space, and x t The hidden representation of each variable in :

[0159]

[0160] Among them, tanh is the activation function, yes The hidden representation of k = 1, 2, ..., n, is the embedding vector, is the bias parameter, which is used to increase the flexibility of the function and improve the fitting ability of the neuron;

[0161] The second layer of the feed-forward neural network is for each hidden representation Assign weights:

[0162]

[0163] in, It is hidden representation The weight of is a variable-order context vector, the vector v is randomly initialized and learned during the training process;

[0164] based on and x t Convert to spatial semantic sequence

[0165]

[0166] Use LSTM module to classify spatial semantic sequences The temporal dependency of As the input of the t-th LSTM unit, the hidden state is output by the t-th LSTM unit

[0167]

[0168] in, is the hidden state output by the t-1th LSTM unit, is the hidden state output by the t-th LSTM unit, i t is the output of the input gate of the t-th LSTM unit, f t is the output of the forget gate of the t-th LSTM unit, o t is the output of the output gate of the t-th LSTM unit, σ represents the sigmoid activation function, and is the bias parameter, ⊙ is the element-wise multiplication, W i , W f , W o and W c is the cycle weight, U i , U f , U o and U c is the input weight, c t is the state of the t-th LSTM unit, c t-1 is the state of the t-1th LSTM unit, is an intermediate variable.

[0169] The other steps and parameters are the same as those in Specific Embodiments 1 to 7.

[0170] The variable distillation attention module of this embodiment limits the contribution of irrelevant variables to spatial semantics by assigning smaller weights to irrelevant variables.

[0171] Specific implementation method 9: This implementation method is different from any one of specific implementation methods 1 to 8 in that the working process of the T-net branch is as follows:

[0172] Step 1: Hide the state and the target variable sequence y t As the input of the tth knowledge-enhanced LSTM unit in the encoder module, that is, but

[0173]

[0174] in, is the hidden state output by the t-1th knowledge-enhanced LSTM unit in the encoder module, is the candidate state of the tth knowledge-enhanced LSTM unit in the encoder module;

[0175] According to the hidden state and Generate the hidden state of the tth knowledge-enhanced LSTM unit:

[0176]

[0177] in, represents the transpose of a matrix, is a learnable weight parameter, is the auxiliary context vector, s t yes The importance score of t yes The weight of

[0178] Finally, the gated fusion unit is used to generate the hidden state When generating the current hidden state, the gated fusion unit weighs the network from the candidate states How much new information is considered by the auxiliary context vector ac can be expressed as follows:

[0179]

[0180] Among them, W a and W h is a learnable weight parameter, is the hidden state output by the tth knowledge-enhanced LSTM unit in the encoder module; g is converted to t The value of is mapped to the interval [0, 1], when g t = 0, the gated fusion will ignore the information of ac;

[0181] Step 2: As the input of each knowledge-enhanced LSTM unit in the decoder module, for the τth knowledge-enhanced LSTM unit in the decoder module, based on the temporal attention mechanism, and the hidden state d of the τ-1th knowledge-enhanced LSTM unit in the decoder module τ-1 Generate context vector hc τ :

[0182]

[0183] in, and is a learnable weight parameter, ⊙ is an element-wise multiplication, e t yes The importance score, β t yes The attention weight of

[0184] According to β tright Perform weighted summation to generate context vector hc τ :

[0185]

[0186] Step 3: Pass the hidden state of the τ-1th knowledge-enhanced LSTM unit through the multilayer perceptron, output the prediction result of the τ-1th knowledge-enhanced LSTM unit through the multilayer perceptron, and compare the prediction result of the τ-1th knowledge-enhanced LSTM unit with the context vector hc τ Perform cascade operations;

[0187] The cascade operation result and hidden state d τ-1 and hidden state As the input of the τth knowledge-enhanced LSTM unit, that is,

[0188] Among them, [:] is a cascade operation, is the prediction result of the τ-1th knowledge-enhanced LSTM unit in the decoder module, d τ is the hidden state of the τth knowledge-enhanced LSTM unit in the decoder module;

[0189] The hidden state d τ After passing through a multi-layer perceptron, the prediction result of the τth knowledge-enhanced LSTM unit is output through the multi-layer perceptron.

[0190] The other steps and parameters are the same as those in Specific Embodiments 1 to 8.

[0191] The knowledge-enhanced LSTM is a variant of the standard LSTM. Compared with the standard LSTM, it adds an interactive attention module (IA) and a gated fusion. The output of the standard LSTM is converted into a candidate state using the symbol and Represent the candidate state and hidden state of KeLSTM at time t respectively. The output of the NP-net branch As auxiliary information of the encoder to enhance the representation of the target variable. At time t, KeLSTM transforms the current target variable and the previous hidden state into candidate states through its three gate units. After that, KeLSTM fuses the auxiliary information matrix To enhance the candidate state, the hidden state is obtained according to the candidate state. The temporal attention mechanism selects the hidden state of the encoder output according to its importance to the decoder.

[0192] Specific implementation method 10: This implementation method is different from any one of the specific implementation methods 1 to 9 in that the hidden state d τAfter passing through the multi-layer perceptron, the prediction result of the τth knowledge-enhanced LSTM unit is output by the multi-layer perceptron, specifically:

[0193]

[0194] in, is a learnable weight parameter, is the prediction result of the τth knowledge-enhanced LSTM unit.

[0195] The other steps and parameters are the same as those in Specific Embodiments 1 to 9.

[0196] The present invention selects part of the data from the public data set of the OpenSky website to construct an experimental data set. The public data set contains rich flight information, usually including the following main parts: aircraft identification (including 24-bit ICAO address, flight number, airline information, etc.), location data (including aircraft latitude and longitude, pressure altitude, geometric altitude, heading, etc.), motion information (aircraft ground speed, vertical speed, etc.), timestamp (the time of each data record, usually in Unix timestamp format, used to identify the collection time point of the data), transponder information (including transponder code, whether there is special positioning information, etc.). Some trajectory data are selected as the experimental data set, and the training set and test set are divided according to the ratio of 8:2. Some variables contained in the data set and their meanings are shown in Table 1:

[0197] Table 1 Specific data of OpenSky dataset

[0198]

[0199] The above calculation examples of the present invention are only used to explain the calculation model and calculation process of the present invention in detail, and are not intended to limit the implementation methods of the present invention. For ordinary technicians in the relevant field, other different forms of changes or modifications can be made based on the above description. It is impossible to list all the implementation methods here. All obvious changes or modifications derived from the technical solution of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for predicting aircraft trajectories in multi-dimensional variable scenarios based on multi-attention collaborative networks, characterized in that: The method specifically comprises the following steps: Step 1: Collect M sets of trajectory data of the aircraft, and each set of trajectory data includes time data, speed data, position data and heading angle data; After preprocessing each set of collected trajectory data respectively, the position data in each set of preprocessed trajectory data is reconstructed to obtain each set of processed trajectory data; Step 2: cluster each group of processed trajectory data to obtain each type of trajectory data; Step 3: Build a prediction model based on multi-attention collaborative network for each type of trajectory data. The prediction model includes two branches, T-net and NP-net, and use each type of trajectory data as the training set of the corresponding prediction model. After training each prediction model using the training set, each trained prediction model is obtained; Step 4: Use the data to be predicted as the input of the corresponding trained prediction model to obtain the trajectory prediction result.

2. According to claim 1, a method for predicting aircraft trajectory in multi-dimensional variable scenarios based on multi-attention collaborative network is characterized in that: The preprocessing of each group of collected trajectory data is performed separately, specifically: Invalid values ​​in the speed data, position data and heading angle data are removed, and noise in the speed data, position data and heading angle data is filtered.

3. The method for predicting aircraft trajectory in multi-dimensional variable scenarios based on multi-attention collaborative network according to claim 2 is characterized in that: The reconstructing of the position data in each set of preprocessed trajectory data is specifically as follows: Take the position data in the m'th group of trajectory data as an example: Step 11: Set the distance threshold; Step 12: Calculate the distance between the position data corresponding to two adjacent time points; If the distance between the position data corresponding to two adjacent time points is greater than the set distance threshold, the two adjacent time points are marked as discontinuity points; If the distance between the position data corresponding to two adjacent time points is less than or equal to the set distance threshold, no processing is required; After respectively judging the position data corresponding to every two adjacent time points in the m'th group of trajectory data, all discontinuity points in the position data of the m'th group of trajectory are obtained; Step 13, respectively obtain the number of data points between every two adjacent discontinuity points, and compare the obtained number of data points with the trajectory noise threshold; If the number of data points between two adjacent discontinuity points is greater than the trajectory noise threshold, the trajectory data corresponding to the two adjacent discontinuity points and the data points between the two adjacent discontinuity points are saved as a new set of trajectory data; If the number of data points between two adjacent discontinuity points is less than or equal to the trajectory noise threshold, the trajectory data corresponding to the two adjacent discontinuity points and the data points between the two adjacent discontinuity points are regarded as noise removal; Step 14: remove the new trajectory data saved in step 13 and the noise data removed in step 13 from the m'th group of trajectory data, and use the remaining data as the updated m'th group of trajectory data; After the position data in each set of trajectory data are processed respectively, the updated trajectory data and the saved new trajectory data are used as each set of processed trajectory data.

4. The method for predicting aircraft trajectory in multi-dimensional variable scenarios based on multi-attention collaborative network according to claim 3 is characterized in that: The clustering of each group of processed trajectory data is performed based on the position data in each group of trajectory data. The clustering process is: Step 21: Calculate the distance matrix D based on the position data in each set of processed trajectory data: Among them, the trajectory set E t is a set of processed trajectory data. tj′ is the set of trajectories E t The j′th group of trajectory data in E, j′=1,2,…,N, N is the set E t The number of groups of trajectory data in It is the set E t The similarity between the first group of trajectory data and the j′th group of trajectory data; Step 22: Initialize the number of cluster centers i′=1; Step 2: According to the distance matrix D, the trajectory set E t Each group of trajectory data in is clustered into i′ class; Step 24: Calculate the clustering error SSE according to the clustering result. If the clustering error converges, the final clustering result is obtained. If the clustering error does not converge, set i′=i′+1 and return to execute step 23.

5. The method for predicting aircraft trajectory in multi-dimensional variable scenarios based on multi-attention collaborative network according to claim 4 is characterized in that: The set E t The calculation method of the similarity between the first group of trajectory data and the j′th group of trajectory data is: Among them, h(E t1 ,E tj′ ) is an intermediate variable, a is the position point in the first set of trajectory data, and b is the position point in the j′th set of trajectory data; 6. The method for predicting aircraft trajectory in multi-dimensional variable scenarios based on multi-attention collaborative network according to claim 5 is characterized in that: The calculation process of the clustering error SSE is: Among them, C i′ represents the i′th cluster, and p represents C i′ The position data in any set of trajectory data, m i′ For C i′ The centroid of the position data in all group trajectory data, k′ is the total number of categories.

7. The method for predicting aircraft trajectory in multi-dimensional variable scenarios based on multi-attention collaborative network according to claim 6 is characterized in that: The training process of the prediction model is: Take any prediction model as an example Step 31: Use each set of trajectory data in the training set of the current prediction model to construct a set of target variable sequences and non-prediction variable sequences; Among them, a set of target variable sequences and non-predictor variable sequences are represented as and x t is the non-predictor variable corresponding to the t-th time point in the trajectory data, They are x t The first, second, and nth variables in ; Where T is the number of time points in a set of trajectory data; Step 3.2: Sequence non-predictor variables As the input of the NP-net branch, the hidden state at each time point is output through the NP-net branch Step 33: The T-net branch includes an encoder module and a decoder module, the encoder module includes T knowledge-enhanced LSTM units, and the decoder module includes Δ knowledge-enhanced LSTM units; Output the hidden state of the NP-net branch and the target variable sequence As the input of the encoder module, the target variable sequence is output through the encoder module The hidden state sequence of Will hide the sequence As the input of the decoder module, the decoder module outputs the hidden state based on the temporal attention mechanism, passes the hidden state output by the decoder module through the multi-layer perceptron, and outputs the prediction result through the multi-layer perceptron; Step 34: Calculate the loss function according to the prediction result output in step 33, adjust the parameters of the prediction model based on the loss function calculation result and the back propagation learning algorithm, and stop training when the set maximum number of training times is reached to obtain a trained prediction model.

8. The method for predicting aircraft trajectory in multi-dimensional variable scenarios based on multi-attention collaborative network according to claim 7 is characterized in that: The NP-net branch includes a variable distillation attention module and an LSTM module, wherein the variable distillation attention module includes T variable distillation attention units, and the LSTM module includes T LSTM units; The working process of the NP-net branch is as follows: Each variable distillation attention unit includes two layers of forward neural networks. The first layer of the forward neural network of the t-th variable distillation attention unit converts x t Each variable in is mapped to the m-dimensional space, and x t The hidden representation of each variable in : Among them, tanh is the activation function, yes The hidden representation of k = 1, 2, ..., n, is the embedding vector, is the bias parameter; The second layer of the feed-forward neural network is for each hidden representation x t ' k Assign weights: in, is the hidden representation x t ' k The weight of is a variable-order context vector; based on and x t Convert to spatial semantic sequence Spatial semantic sequence As the input of the t-th LSTM unit, the hidden state is output by the t-th LSTM unit in, is the hidden state output by the t-1th LSTM unit, is the hidden state output by the t-th LSTM unit, i t is the output of the input gate of the t-th LSTM unit, f t is the output of the forget gate of the t-th LSTM unit, o t is the output of the output gate of the t-th LSTM unit, σ represents the sigmoid activation function, and is the bias parameter, ⊙ is the element-wise multiplication, W i , W f , W o and W c is the cycle weight, U i , U f , U o and U c is the input weight, c t is the state of the t-th LSTM unit, c t-1 is the state of the t-1th LSTM unit, is an intermediate variable.

9. The method for predicting aircraft trajectory in multi-dimensional variable scenarios based on multi-attention collaborative network according to claim 8 is characterized in that: The working process of the T-net branch is as follows: Step 1: Hide the state and the target variable sequence y t As the input of the tth knowledge-enhanced LSTM unit in the encoder module, that is, but in, is the hidden state output by the t-1th knowledge-enhanced LSTM unit in the encoder module, h′ t y is the candidate state of the tth knowledge-enhanced LSTM unit in the encoder module; According to the hidden state and h′ t y Generate the hidden state of the tth knowledge-enhanced LSTM unit: in, represents the transpose of a matrix, is the weight parameter, is the auxiliary context vector, s t yes The importance score of t yes The weight of g t =σ(W a ac+W h h t ′ y ) Among them, W a and W h is the weight parameter, is the hidden state output by the tth knowledge-enhanced LSTM unit in the encoder module; Step 2: As the input of each knowledge-enhanced LSTM unit in the decoder module, for the τth knowledge-enhanced LSTM unit in the decoder module, based on the temporal attention mechanism, and the hidden state d of the τ-1th knowledge-enhanced LSTM unit in the decoder module τ-1 Generate context vector hc τ : in, and is the weight parameter, ⊙ is the element-wise multiplication, e t yes The importance score, β t yes The attention weight of According to β t right Perform weighted summation to generate context vector hc τ : Step 3: Pass the hidden state of the τ-1th knowledge-enhanced LSTM unit through the multilayer perceptron, output the prediction result of the τ-1th knowledge-enhanced LSTM unit through the multilayer perceptron, and compare the prediction result of the τ-1th knowledge-enhanced LSTM unit with the context vector hc τ Perform cascade operations; The cascade operation result and hidden state d τ-1 and hidden state As the input of the τth knowledge-enhanced LSTM unit, that is, Among them, [:] is a cascade operation, is the prediction result of the τ-1th knowledge-enhanced LSTM unit in the decoder module, d τ is the hidden state of the τth knowledge-enhanced LSTM unit in the decoder module; Hide the state d τ After passing through a multi-layer perceptron, the prediction result of the τth knowledge-enhanced LSTM unit is output through the multi-layer perceptron.

10. The method for predicting aircraft trajectory in multi-dimensional variable scenarios based on multi-attention collaborative network according to claim 9 is characterized in that: The hidden state d τ After passing through the multi-layer perceptron, the prediction result of the τth knowledge-enhanced LSTM unit is output by the multi-layer perceptron, specifically: in, is the weight parameter, is the prediction result of the τth knowledge-enhanced LSTM unit.

Images