An aircraft trajectory prediction method in a multi-dimensional variable scene based on a multi-attention collaborative network
By employing a multi-attention collaborative network approach, utilizing T-net and NP-net branching and variable distillation attention and knowledge-enhanced LSTM networks, the accuracy problem of aircraft trajectory prediction in multivariate scenarios is solved, achieving efficient prediction in complex environments.
Patent Information
- Application Number
- CN202411321178.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-23
- Publication Date
- 2026-02-06
- Estimated Expiration
- 2044-09-23
AI Technical Summary
Existing data-driven aircraft trajectory prediction methods struggle to adequately model the complex relationships between variables in multivariate scenarios, resulting in poor prediction accuracy and an inability to effectively address aircraft motion in complex nonlinear environments.
We employ a multi-attention collaborative network-based approach, which involves data collection, preprocessing, clustering, and constructing a multi-attention collaborative network prediction model. We utilize T-net and NP-net branches for trajectory prediction and combine variable distillation attention with a knowledge-enhanced LSTM network to capture the interactions and temporal dependencies between variables.
It improves the accuracy and adaptability of aircraft trajectory prediction, effectively captures the potential interaction between target and non-predictable variables in complex nonlinear environments, reduces the negative impact of irrelevant variables, and improves prediction accuracy.
Smart Images
Figure CN119204325B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of trajectory prediction technology, specifically relating to a method for predicting aircraft trajectories in multidimensional variable scenarios based on multi-attention collaborative networks. Background Technology
[0002] Simulated entities are the basic functional units constituting a distributed simulation system. The main purpose of distributed simulation is to construct a realistic and perceptually consistent virtual environment among the participating simulated entities. When the state of an entity in the experiment changes or its position in the virtual environment changes, all participating entities can promptly detect the update of the entity's state. However, due to network communication capacity limitations and uncertain delays in information transmission, different simulated entities often cannot achieve consistency in their observation of the same entity's state update, leading to spatiotemporal inconsistencies in the distributed simulation process. As distributed simulation systems develop towards larger scales, more entities, wider geographical areas, more diverse structures, and longer durations, the level of detail in the simulation process is also increasing, making the spatiotemporal inconsistency problem in distributed simulation systems more prominent. Large-scale distributed simulation systems have high requirements for spatiotemporal consistency. The existence of spatiotemporal inconsistencies can lead to inconsistencies in perception among simulated entities, resulting in adverse effects such as reversed causality and misunderstandings, undermining the realism and effectiveness of the simulation, and affecting the credibility of the simulation results.
[0003] To address the spatiotemporal inconsistency issue, the inference and localization algorithm utilizes prediction techniques between entities to compensate for spatiotemporal inconsistencies caused by network transmission instability, thereby ensuring the realism of the simulation and the reliability of the results. However, it relies solely on short-period historical data for entity state prediction, making it difficult to mine the motion characteristics of various highly maneuverable targets from both temporal and spatial dimensions. It also cannot handle abrupt changes in the predicted object's motion behavior, affecting the accuracy of entity state prediction. Therefore, other more efficient and accurate trajectory prediction methods are needed to predict the trajectories of simulated entities such as aircraft in distributed simulation systems.
[0004] Most existing aircraft trajectory prediction methods are based on predefined motion models, using algorithms such as Markov chains and Kalman filters for trajectory prediction. However, when performing missions, aircraft often operate in complex nonlinear environments. Their motion behavior is not only constrained by the aircraft's own dynamics but also influenced by various external factors such as weather conditions, terrain, and environmental disturbances. These factors make the aircraft's motion state difficult to predict. Traditional physical models struggle to accurately describe the aircraft's true motion behavior in such complex environments, leading to decreased prediction accuracy. Furthermore, the construction of physical models typically relies on numerous assumptions and simplifications; when these assumptions no longer hold true, the model's applicability and accuracy also decrease. In recent years, with the development of machine learning and deep learning methods, using data-driven approaches to learn the complex motion patterns of aircraft to improve trajectory prediction accuracy has become feasible. Therefore, trajectory tracking and prediction algorithms based on deep neural networks have attracted increasing attention from researchers.
[0005] Data-driven trajectory prediction methods rely on a large amount of historical and real-time data to predict aircraft motion, offering significant advantages in handling complex nonlinear environments. By learning from vast amounts of historical data, data-driven methods do not require in-depth knowledge of the aircraft's physical characteristics, which reduces modeling difficulty and improves model adaptability, effectively addressing diverse flight mission scenarios. However, data-driven methods still have significant limitations in multivariate prediction. When dealing with multivariate data, these methods struggle to fully model the complex relationships between variables, often requiring complex feature selection; otherwise, information loss or noise introduction can easily occur, leading to poor prediction accuracy.
[0006] Therefore, developing a spacecraft trajectory prediction technology under multi-dimensional variable scenarios is of great significance for ensuring anti-interference performance, improving the spatiotemporal consistency of spacecraft state perception by various resources during virtual-real collaborative testing, and reducing the impact of spacecraft state perception lag on the realism of complex environment construction. Summary of the Invention
[0007] The purpose of this invention is to address the problem that existing data-driven prediction methods in multivariate scenarios struggle to adequately model the complex relationships between variables, leading to poor prediction accuracy. To address this, a method for predicting aircraft trajectories in multivariate scenarios based on multi-attention collaborative networks is proposed.
[0008] The technical solution adopted by this invention to solve the above-mentioned technical problems is: a method for predicting aircraft trajectory in multi-dimensional variable scenarios based on multi-attention collaborative networks, the method specifically including the following steps:
[0009] Step 1: Collect M sets of trajectory data from the aircraft, with each set including time data, velocity data, position data, and heading angle data;
[0010] After preprocessing each set of collected trajectory data, the position data in each set of preprocessed trajectory data is reconstructed to obtain the processed sets of trajectory data.
[0011] Step 2: Cluster the processed trajectory data to obtain trajectory data of different categories;
[0012] Step 3: Construct a prediction model based on a multi-attention collaborative network for each type of trajectory data. The prediction model includes two branches: T-net and NP-net. Each type of trajectory data is used as the training set for the corresponding prediction model.
[0013] After training each prediction model using the training set, we obtain the trained prediction models.
[0014] Step 4: Use the data to be predicted as input to the corresponding trained prediction model to obtain the trajectory prediction result.
[0015] Furthermore, the preprocessing of each set of collected trajectory data is specifically as follows:
[0016] Remove invalid values from the speed, position, and heading data, and filter out noise from the speed, position, and heading data.
[0017] Furthermore, the reconstruction of the position data in each group of preprocessed trajectory data specifically involves:
[0018] Taking the position data in the m'th group of trajectory data as an example:
[0019] Step 11: Set the distance threshold;
[0020] Step 1 and Step 2: Calculate the distance between the location data corresponding to two adjacent time points;
[0021] If the distance between the location data corresponding to two adjacent time points is greater than the set distance threshold, the two adjacent time points are marked as discontinuities.
[0022] If the distance between the location data corresponding to two adjacent time points is less than or equal to the set distance threshold, no processing is required;
[0023] After judging the position data corresponding to every two adjacent time points in the m'th group of trajectory data, all discontinuities in the position data of the m'th group of trajectory are obtained;
[0024] Step 13: Obtain the number of data points between every two adjacent discontinuities, and compare the obtained number of data points with the trajectory noise threshold.
[0025] If the number of data points between two adjacent discontinuities is greater than the trajectory noise threshold, then the trajectory data corresponding to the two adjacent discontinuities and the data points between the two adjacent discontinuities are saved as a new set of trajectory data.
[0026] If the number of data points between two adjacent discontinuities is less than or equal to the trajectory noise threshold, then the trajectory data corresponding to the two adjacent discontinuities and the data points between the two adjacent discontinuities are considered as noise and are removed.
[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 processing the location data in each group of trajectory data, the updated trajectory data and the saved new trajectory data are used as the processed trajectory data for each group.
[0029] Furthermore, the clustering of the processed trajectory data is based on the location data in each group of trajectory data, and the clustering process is as follows:
[0030] Step 2.1: Calculate the distance matrix D based on the position data in each group of processed trajectory data.
[0031]
[0032] Among them, the trajectory set E t E is a collection of processed trajectory data. tj′ It is the set of trajectories E t The j′-th group of trajectory data in the dataset, j′=1,2,…,N, where N is a set E t The number of groups of trajectory data. It is set E t The similarity between the first group of trajectory data and the j′ group of trajectory data;
[0033] Step 22: Initialize the number of cluster centers i′ = 1;
[0034] Steps 2 and 3: Based on the distance matrix D, set the trajectory E t The trajectory data in the data are clustered into class i′;
[0035] Step 24: Calculate the clustering error SSE based on the clustering results. If the clustering error converges, the final clustering result is obtained. If the clustering error does not converge, let i′ = i′ + 1 and return to step 23.
[0036] Furthermore, the set E t The method for calculating the similarity between the first group of trajectory data and the j′ group of trajectory data is as follows:
[0037]
[0038] Where h(E) t1 E tj′ ) are intermediate variables, a is the position point in the first group of trajectory data, and b is the position point in the j′ group of trajectory data;
[0039]
[0040] Furthermore, the calculation process for the clustering error SSE is as follows:
[0041]
[0042] Among them, C i′ Let p represent the i′-th cluster, and p represent C. i′ The position data in any set of trajectory data, m i′ Belongs to C i′ The centroid of the location data in all group trajectory data, k′ is the total number of categories.
[0043] Furthermore, the training process of the prediction model is as follows:
[0044] Take any prediction model as an example
[0045] Step 3: 1. Construct a set of target variable sequences and non-predictor variable sequences using each set of trajectory data in the training set of the current prediction model;
[0046] Here, a set of target variable sequences and non-predictor variable sequences are respectively represented as and x t It is the non-predictable variable corresponding to the t-th time point in the trajectory data. They are x t The first, second, and nth variables in the list;
[0047] Where T is the number of time points in a set of trajectory data;
[0048] Step 3.2: Transform the non-predictor variable sequence As input to the NP-net branch, the hidden state at each time point is output through the NP-net branch.
[0049] Step 3: 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] Hidden state of NP-net branch output and target variable sequence As input to the encoder module, the encoder module outputs a sequence of target variables. Hidden state sequence
[0051] Hidden representation sequence As input to the decoder module, the decoder module outputs the hidden state based on the temporal attention mechanism. The hidden state output by the decoder module is then passed through a multilayer perceptron, and the prediction result is output through the multilayer perceptron.
[0052] Step 3-4: Calculate the loss function based on the prediction results output in Step 3-3. Adjust the parameters of the prediction model based on the loss function calculation results and the backpropagation learning algorithm until the maximum number of training iterations is reached, then stop training 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 specific working process of the NP-net branch is as follows:
[0055] Each variable distillation attention unit includes two layers of feedforward neural networks. The first layer of the feedforward neural network in the t-th variable distillation attention unit will input x... t Each variable in the equation is mapped to an m-dimensional space to obtain x. t Hidden representation of each variable in the code:
[0056]
[0057] Where tanh is the activation function. yes The hidden representation, k = 1, 2, ..., n, It is an embedding vector. It is a bias parameter;
[0058] The second feedforward neural network represents each hidden representation. Assign weights:
[0059]
[0060] in, It is a hidden representation The weight, It is a variable-order context vector;
[0061] based on and x t Convert to spatial semantic sequence
[0062]
[0063] spatial semantic sequence The hidden state is input to the t-th LSTM unit and output through the t-th LSTM unit.
[0064]
[0065] in, It is the hidden state output by the (t-1)th LSTM unit. It is the hidden state output by the t-th LSTM unit, i t f is the output of the input gate of the t-th LSTM unit. t It is the output of the forget gate of the t-th LSTM unit, o t It is the output of the output gate of the t-th LSTM unit, and σ represents the sigmoid activation function. and It is the bias parameter, ⊙ is element-wise multiplication, W i W f W o and W c It is the recurrent weight, U i U f U o and U c It is the input weight, c t It is the state of the t-th LSTM unit, c t-1 This is the state of the (t-1)th LSTM unit. It 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 to the t-th knowledge-enhanced LSTM unit in the encoder module, i.e. but
[0068]
[0069] in, It is the hidden state output by the (t-1)th knowledge-enhanced LSTM unit in the encoder module. It is a candidate state of the t-th knowledge-enhanced LSTM unit in the encoder module;
[0070] Based on the hidden state and Generate the hidden state of the t-th knowledge-enhanced LSTM unit:
[0071]
[0072] in, Represents the transpose of a matrix. These are weight parameters. It is an auxiliary context vector, s t yes Importance score, γ t yes The weights;
[0073]
[0074] Among them, W a and W h These are weight parameters. It is the hidden state output by the t-th knowledge-enhanced LSTM unit in the encoder module;
[0075] Step 2, As the input to each knowledge-enhanced LSTM unit in the decoder module, for the τ-th knowledge-enhanced LSTM unit in the decoder module, based on the time attention mechanism, The hidden state d of the (τ-1)th knowledge-enhanced LSTM unit in the decoder module τ-1 Generate context vector hc τ :
[0076]
[0077] in, and These are weight parameters, ⊙ represents element-wise multiplication, and e t yes Importance score, β t yes Attention weights;
[0078] According to β t right Perform weighted summation to generate the context vector hc τ :
[0079]
[0080] Step 3: Pass the hidden state of the (τ-1)th knowledge-enhanced LSTM unit through a multilayer perceptron. The prediction result of the (τ-1)th knowledge-enhanced LSTM unit is then output through the multilayer perceptron. Finally, the prediction result of the (τ-1)th knowledge-enhanced LSTM unit is compared with the context vector hc. τ Perform cascading operations;
[0081] The result of the cascade operation and the hidden state d τ-1 and hidden state As the input to the τth knowledge-enhanced LSTM unit, i.e.
[0082] Here, [:] represents a cascading operation. It is the prediction result of the (τ-1)th knowledge-enhanced LSTM unit in the decoder module, d τ It is the hidden state of the τth knowledge-enhanced LSTM unit in the decoder module;
[0083] Hidden state d τ After passing through a multilayer perceptron, the prediction result of the τth knowledge-enhanced LSTM unit is output.
[0084] Furthermore, the hidden state d τ After passing through a multilayer perceptron, the prediction result of the τth knowledge-enhanced LSTM unit is output, specifically:
[0085]
[0086] in, These are weight parameters. This is the prediction result of the τth knowledge-enhanced LSTM unit.
[0087] The beneficial effects of this invention are:
[0088] This invention first designs a variable distillation attention mechanism as the first layer network. This network weakens the negative impact of unpredictable variables and transforms the current sequence into spatial semantics. Furthermore, a novel Long Short-Term Memory (LSTM) network variant, KeLSTM, is designed to compute interactions between variables, mitigating information conflict and distinguishing the temporal impact of variables on the task. This method reduces the negative impact of irrelevant variables, emphasizes the contribution of important variables to prediction, and successfully captures the potential interactions between past target variables, unpredictable variables, and the upcoming target, particularly the subtle differences between the influence of the target and unpredictable variables. This method improves prediction accuracy without requiring complex feature selection and remains highly adaptable to complex nonlinear environments. Attached Figure Description
[0089] Figure 1 This is a flowchart of the predictive model's workflow;
[0090] In this invention, y t As shown in the figure, χ t ,Will As shown in the figure
[0091] Figure 2 This is a flowchart of a method for predicting aircraft trajectory in a multi-dimensional variable scenario based on a multi-attention collaborative network, according to the present invention. Detailed Implementation
[0092] Specific implementation method one: Combining Figure 2 This embodiment describes a method for predicting aircraft trajectories in a multi-dimensional variable scenario based on a multi-attention collaborative network. The method specifically includes the following steps:
[0093] Step 1: Collect M sets of trajectory data from the aircraft, with each set including time data, velocity data, position data, and heading angle data;
[0094] After preprocessing each set of collected trajectory data, the position data in each set of preprocessed trajectory data is reconstructed to obtain the processed sets of trajectory data.
[0095] Step 2: Cluster the processed trajectory data (clustering methods include but are not limited to the k-medoids algorithm) to obtain trajectory data of each class; each class of trajectory data corresponds to a trajectory route, the trajectory change trend in each class is similar, and the trajectory change trend between different classes is different.
[0096] Step 3: Construct a prediction model based on a multi-attention collaborative network (i.e., a MACN network) for each type of trajectory data. The prediction model includes two branches: T-net and NP-net, and each type of trajectory data is used as the training set for the corresponding prediction model.
[0097] After training each prediction model using the training set, we obtain the trained prediction models.
[0098] Step 4: Use the data to be predicted as input to the corresponding trained prediction model (the trajectory and route category of the data to be predicted is known, so the data to be predicted can be used as input to the prediction model corresponding to the known trajectory and route category) to obtain the trajectory prediction result.
[0099] The NP-net branch of this invention focuses primarily on non-predictable variables, capturing the most relevant variables and time dependencies; the T-net branch predicts the target variable, connecting both the encoder and decoder modules to the NP-net branch, thereby utilizing the output of the NP-net branch to assist learning and decision-making; knowledge-enhanced LSTM (KeLSTM) is used as the encoder and decoder modules of the T-net branch, selectively enhancing the latent representation of the target variable, handling conflicts between the target variable and non-predictable variables, and improving the accuracy of prediction.
[0100] Specific Implementation Method Two: This implementation method differs from Specific Implementation Method One in that the preprocessing of each set of collected trajectory data is performed separately, specifically as follows:
[0101] Remove invalid values from the speed, position, and heading data, and filter out noise from the speed, position, and heading data.
[0102] The other steps and parameters are the same as in Specific Implementation Method 1.
[0103] Specific Implementation Method Three: This implementation method differs from Specific Implementation Method One or Two in that the reconstruction of the position data in each group of preprocessed trajectory data specifically involves:
[0104] Taking the position data in the m'th group of trajectory data as an example:
[0105] Step 11: Set the distance threshold;
[0106] Step 1 and Step 2: Calculate the distance between the location data corresponding to two adjacent time points;
[0107] If the distance between the location data corresponding to two adjacent time points is greater than the set distance threshold, the two adjacent time points are marked as discontinuities.
[0108] If the distance between the location data corresponding to two adjacent time points is less than or equal to the set distance threshold, no processing is required;
[0109] After judging the position data corresponding to every two adjacent time points in the m'th group of trajectory data, all discontinuities in the position data of the m'th group of trajectory are obtained;
[0110] Step 13: Obtain the number of data points between every two adjacent discontinuities, and compare the obtained number of data points with the trajectory noise threshold.
[0111] If the number of data points between two adjacent discontinuities is greater than the trajectory noise threshold, then the trajectory data corresponding to the two adjacent discontinuities and the data points between the two adjacent discontinuities are saved as a new set of trajectory data.
[0112] If the number of data points between two adjacent discontinuities is less than or equal to the trajectory noise threshold, then the trajectory data corresponding to the two adjacent discontinuities and the data points between the two adjacent discontinuities are considered as noise and are removed.
[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 processing the location data in each group of trajectory data, the updated trajectory data and the saved new trajectory data are used as the processed trajectory data for each group.
[0115] Other steps and parameters are the same as in specific implementation method one or two.
[0116] The distance threshold and trajectory noise threshold can be selected according to the actual situation.
[0117] Specific Implementation Method Four: This implementation method differs from Specific Implementation Methods One to Three in that the clustering of the processed trajectory data is based on the location data within each group of trajectory data. The clustering process is as follows:
[0118] Step 2.1: Calculate the distance matrix D based on the position data in each group of processed trajectory data.
[0119]
[0120] Among them, the trajectory set E t E is a collection of processed trajectory data. tj′ It is the set of trajectories E t The j′-th group of trajectory data in the dataset, j′=1,2,…,N, where N is a set E t The number of groups of trajectory data. It is set E t The similarity between the first group of trajectory data and the j′ group of trajectory data;
[0121] Step 22: Initialize the number of cluster centers i′ = 1;
[0122] Steps 2 and 3: Based on the distance matrix D, set the trajectory E t The trajectory data in the data are clustered into class i′;
[0123] Step 24: Calculate the clustering error SSE based on the clustering results. If the clustering error converges (convergence judgment is only performed when i′=2, and if i′=1, it is directly treated as non-convergence), the final clustering result is obtained. If the clustering error does not converge, let i′=i′+1 and return to step 23.
[0124] The other steps and parameters are the same as those in one of the specific implementation methods one to three.
[0125] Clustering results are typically evaluated using the silhouette coefficient, which measures the quality of clustering from the perspectives of intra-cluster compactness and inter-cluster separation. This yields a criterion for evaluating the quality of clustering. The silhouette coefficient can be expressed as follows:
[0126]
[0127] Here, 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 density between the clusters to which x belongs; the smaller the value, the more compact the clustering. The average distance from x to all trajectories in another cluster is calculated, and similarly, the average distance for each other cluster is calculated. The smallest average distance is taken as b(x); the larger this value, the better the separation. The closer the silhouette coefficient is to 1, the more ideal the clustering effect.
[0128] Specific Implementation Method Five: This implementation method differs from Specific Implementation Methods One to Four in that the set E... t The method for calculating the similarity between the first group of trajectory data and the j′ group of trajectory data is as follows:
[0129]
[0130] Among them, h(E) t1 E tj′ (E) is an intermediate variable, where a is a position point in the first group of trajectory data and b is a position point in the j′ group of trajectory data. First, the distance between position point a in the first group of trajectory data and each position point b in the j′ group of trajectory data is calculated. Then, the minimum distance corresponding to position point a is selected. Similarly, the minimum distance corresponding to each position point in the first group of trajectory data is calculated. The maximum value among 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 one of the specific implementation methods one to four.
[0135] Specific Implementation Method Six: This implementation method differs from Specific Implementation Methods One to Five in that the calculation process of the clustering error SSE is as follows:
[0136]
[0137] Among them, C i′ Let p represent the i′-th cluster, and p represent C. i′ The position data in any set of trajectory data, m i′ Belongs to C i′ The centroid of the location 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 one of the specific implementation methods one to five.
[0139] Specific implementation method seven: Combining Figure 1 This embodiment is described below. The difference between this embodiment and one of the specific embodiments one through six is that the training process of the prediction model is as follows:
[0140] Take any prediction model as an example
[0141] Step 3: 1. Construct a set of target variable sequences and non-predictor variable sequences using each set of trajectory data in the training set of the current prediction model;
[0142] Here, a set of target variable sequences and non-predictor variable sequences are respectively represented as and x t It is the non-predictable variable corresponding to the t-th time point in the trajectory data. They are x t The first, second, and nth variables in the list;
[0143] Where T is the number of time points in a set of trajectory data;
[0144] For example, when it is necessary to predict location data, the target variable sequence is a sequence of location data in the trajectory data, and the non-predictable variable sequence is a sequence of other data in the trajectory data besides location. That is, only one type of data can be predicted at a time. The target variable sequence only includes the data that needs to be predicted, and the non-predictable 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 separately for each type of data and used for model training.
[0145] Step 3.2: Transform the non-predictor variable sequence As input to the NP-net branch, the hidden state at each time point is output through the NP-net branch.
[0146] Step 3: 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] Hidden state of NP-net branch output and target variable sequence As input to the encoder module, the encoder module outputs a sequence of target variables. Hidden state sequence
[0148] Hidden representation sequence As input to the decoder module, the decoder module outputs the hidden state based on the temporal attention mechanism. The hidden state output by the decoder module is then passed through a multilayer perceptron, and the prediction result is output through the multilayer perceptron.
[0149] Step 3-4: Calculate the loss function based on the prediction results output in Step 3-3. Adjust the parameters of the prediction model based on the loss function calculation results and the backpropagation learning algorithm until the maximum number of training iterations is reached, then stop training to obtain a trained prediction model.
[0150] The other steps and parameters are the same as those in one of the specific implementation methods one to six.
[0151] KeLSTM first computes the unpredicted sequence and target variable y t Interactive attention between them. Then, by selectively fusing the interactive information using a gated fusion unit, the interaction information is enhanced. t Potential representations KeLSTM uses interactive attention to capture important latent representations. For prediction, a gated fusion unit is used to balance the conflict between the target variable and the unpredictable variables. Finally, the prediction model utilizes a multilayer perceptron as the task learning layer (TLL), which will process the state d... τ Convert to predicted value
[0152] The prediction model of this invention uses the root mean square error (RMSE) as the loss function and employs the Adam optimizer to optimize the network parameters. The learning rate is initialized to 0.001 and is reduced according to the training loss. The formula for calculating RMSE can be expressed as follows:
[0153]
[0154] Where Δ is the size of the prediction range, and N is the number of samples. It is the actual value. That is the corresponding predicted value.
[0155] The trained prediction model is then tested.
[0156] Specific Implementation Method Eight: This implementation method differs from one of the specific implementation methods 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 specific working process of the NP-net branch is as follows:
[0158] Each variable distillation attention unit includes two layers of feedforward neural networks. The first layer of the feedforward neural network in the t-th variable distillation attention unit will input x... t Each variable in the equation is mapped to an m-dimensional space to obtain x. t Hidden representation of each variable in the code:
[0159]
[0160] Where tanh is the activation function. yes The hidden representation, k = 1, 2, ..., n, It is an embedding vector. These are bias parameters used to increase the flexibility of the function and improve the fitting ability of neurons;
[0161] The second feedforward neural network represents each hidden representation. Assign weights:
[0162]
[0163] in, It is a hidden representation The weight, It is a variable-order context vector, where vector v is randomly initialized and learned during training;
[0164] based on and x t Convert to spatial semantic sequence
[0165]
[0166] Using LSTM modules to process spatial semantic sequences Modeling the time dependency of spatial semantic sequences The hidden state is input to the t-th LSTM unit and output through the t-th LSTM unit.
[0167]
[0168] in, It is the hidden state output by the (t-1)th LSTM unit. It is the hidden state output by the t-th LSTM unit, i t f is the output of the input gate of the t-th LSTM unit. t It is the output of the forget gate of the t-th LSTM unit, o t It is the output of the output gate of the t-th LSTM unit, and σ represents the sigmoid activation function. and It is the bias parameter, ⊙ is element-wise multiplication, W i W f W o and W c It is the recurrent weight, U i U f U o and U c It is the input weight, c t It is the state of the t-th LSTM unit, c t-1 This is the state of the (t-1)th LSTM unit. It is an intermediate variable.
[0169] The other steps and parameters are the same as those in any of the specific implementation methods one to seven.
[0170] The variable distillation attention module in this embodiment limits the contribution of irrelevant variables to spatial semantics by assigning them smaller weights.
[0171] Specific Implementation Method Nine: This implementation method differs from Specific Implementation Methods One through Eight 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 to the t-th knowledge-enhanced LSTM unit in the encoder module, i.e. but
[0173]
[0174] in, It is the hidden state output by the (t-1)th knowledge-enhanced LSTM unit in the encoder module. It is a candidate state of the t-th knowledge-enhanced LSTM unit in the encoder module;
[0175] Based on the hidden state and Generate the hidden state of the t-th knowledge-enhanced LSTM unit:
[0176]
[0177] in, Represents the transpose of a matrix. These are learnable weight parameters. It is an auxiliary context vector, s t yes Importance score, γ t yes The weights;
[0178] Finally, a gated fusion unit is used to generate the hidden state. When the current hidden state is generated, the gated fusion unit trades the network from the candidate state. The amount of new information considered by the auxiliary context vector ac can be represented as follows:
[0179]
[0180] Among them, W a and W h These are learnable weight parameters. It is the hidden state output by the t-th knowledge-enhanced LSTM unit in the encoder module; g is converted by the logistic sigmoid function. t The value of g is mapped to the interval [0, 1], when g t When = 0, gating fusion will ignore the information of AC;
[0181] Step 2, As the input to each knowledge-enhanced LSTM unit in the decoder module, for the τ-th knowledge-enhanced LSTM unit in the decoder module, based on the time attention mechanism, The hidden state d of the (τ-1)th knowledge-enhanced LSTM unit in the decoder module τ-1 Generate context vector hc τ :
[0182]
[0183] in, and These are learnable weight parameters, ⊙ represents element-wise multiplication, and e t yes Importance score, β t yes Attention weights;
[0184] According to β tright Perform weighted summation to generate the context vector hc τ :
[0185]
[0186] Step 3: Pass the hidden state of the (τ-1)th knowledge-enhanced LSTM unit through a multilayer perceptron. The prediction result of the (τ-1)th knowledge-enhanced LSTM unit is then output through the multilayer perceptron. Finally, the prediction result of the (τ-1)th knowledge-enhanced LSTM unit is compared with the context vector hc. τ Perform cascading operations;
[0187] The result of the cascade operation and the hidden state d τ-1 and hidden state As the input to the τth knowledge-enhanced LSTM unit, i.e.
[0188] Here, [:] represents a cascading operation. It is the prediction result of the (τ-1)th knowledge-enhanced LSTM unit in the decoder module, d τ It is the hidden state of the τth knowledge-enhanced LSTM unit in the decoder module;
[0189] Hidden state d τ After passing through a multilayer perceptron, the prediction result of the τth knowledge-enhanced LSTM unit is output.
[0190] The other steps and parameters are the same as those in one of the specific implementation methods one to eight.
[0191] Knowledge-enhanced LSTM is a variant of standard LSTM, adding an interactive attention module (IA) and gated fusion compared to standard LSTM. It transforms the output of standard LSTM into candidate states using symbolic representation. and Let represent the candidate state and hidden state of KeLSTM at time t, respectively. The output of the NP-net branch... The KeLSTM uses auxiliary information from the encoder to enhance the representation of the target variable. At time t, the KeLSTM transforms the current target variable and the previous hidden state into candidate states through its three gate units. Subsequently, the KeLSTM fuses the auxiliary information matrix... The candidate states are enhanced, and the hidden states are derived from these candidate states. The temporal attention mechanism selects the hidden states from the encoder output based on their importance to the decoder.
[0192] Specific Implementation Method Ten: This implementation method differs from Specific Implementation Methods One to Nine in that the hidden state d is... τAfter passing through a multilayer perceptron, the prediction result of the τth knowledge-enhanced LSTM unit is output, specifically:
[0193]
[0194] in, These are learnable weight parameters. This is the prediction result of the τth knowledge-enhanced LSTM unit.
[0195] The other steps and parameters are the same as those in any of the specific implementation methods one to nine.
[0196] This invention constructs an experimental dataset by selecting a portion of data from the publicly available dataset on the OpenSky website. The publicly available dataset contains rich flight information, typically including the following main components: aircraft identifiers (including 24-bit ICAO address, flight number, airline information, etc.), location data (including aircraft latitude and longitude, barometric altitude, geometric altitude, heading, etc.), motion information (aircraft ground speed, vertical speed, etc.), timestamps (the time of each data record, usually in Unix timestamp format, used to identify the data collection time), and transponder information (including transponder code, whether there is special positioning information, etc.). A portion of the trajectory data was selected as the experimental dataset, and the training and test sets were divided in an 8:2 ratio. Some variables in the dataset and their meanings are shown in Table 1.
[0197] Table 1 shows the specific data for the OpenSky dataset.
[0198]
[0199] The above examples of the present invention are merely illustrative of the computational model and process of the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is impossible to exhaustively list all possible implementations here. Any obvious variations or modifications derived from the technical solutions of the present invention are still within the scope of protection of the present invention.
Claims
1. A method for aircraft trajectory prediction in a multi-dimensional variable scenario based on a multi-attention collaborative network, characterized in that, The method specifically includes the following steps: Step one, collecting aircraft's group trajectory data, and each group trajectory data includes time data, speed data, position data and heading angle data; After preprocessing each set of collected trajectory data, the position data in each set of preprocessed trajectory data is reconstructed to obtain the processed sets of trajectory data. Step 2: Cluster the processed trajectory data to obtain trajectory data of different categories; Step 3: Construct a prediction model based on a multi-attention collaborative network for each type of trajectory data. The prediction model includes two branches: T-net and NP-net. Each type of trajectory data is used as the training set for the corresponding prediction model. After training each prediction model using the training set, we obtain the trained prediction models. The training process of the prediction model is as follows: Take any prediction model as an example Step 3:
1. Construct a set of target variable sequences and non-predictor variable sequences using each set of trajectory data in the training set of the current prediction model; wherein a set of target variable sequences and non-predicted variable sequences are represented as and , is the non-predicted variable corresponding to the time point in the trajectory data, , are the 1st, 2nd, nth variable in , respectively; wherein, is the number of time points in the set of trajectory data; Step three two, the non-predicted variable sequence As input to the NP-net branch, the hidden state at each time point is output by the NP-net branch ; Step three, the T-net branch includes an encoder module and a decoder module, the encoder module includes one knowledge enhanced LSTM unit, and the decoder module includes one knowledge enhanced LSTM unit. hidden state of the NP-net branch output and the target variable sequence as input to the encoder module, the target variable sequence is output by the encoder module a sequence of hidden states ; The hidden representation sequence is obtained As an input of the decoder module, the decoder module outputs a hidden state based on a time attention mechanism, and the hidden state output by the decoder module passes through a multi-layer perceptron to output a prediction result. Step 3-4: Calculate the loss function based on the prediction results output in Step 3-3. Adjust the parameters of the prediction model based on the loss function calculation results and the backpropagation learning algorithm until the maximum number of training iterations is reached, then stop training and obtain the trained prediction model. The NP-net branch includes a variable distillation attention module and an LSTM module, wherein the variable distillation attention module includes variable distillation attention units, and the LSTM module includes LSTM units. The specific working process of the NP-net branch is as follows: Each variable distillation attention unit includes two layers of feedforward neural networks, a first layer of feedforward neural networks of the variable distillation attention unit maps each of the variables to an m-dimensional space, resulting in a hidden representation of each of the variables wherein, is an activation function, is a hidden representation of, , is an embedding vector, is a bias parameter; The second layer forward neural network assigns weights to each hidden representation Assigning weights: wherein is a hidden representation of the weights, is a variable order context vector; Based on and , the is converted into a spatial semantic sequence : To sequence spatial semantics As a first input to a first output hidden state : wherein is the hidden state output by the th LSTM unit, is the hidden state output by the th LSTM unit, is the output of the input gate of the th LSTM unit, is the output of the forget gate of the th LSTM unit, is the output of the output gate of the th LSTM unit, denotes the sigmoid activation function, , , and are bias parameters, is an element-wise multiplication, , , and are recurrent weights, , , and are input weights, is the state of the th LSTM unit, is the state of the th LSTM unit, is an intermediate variable; Step 4: Use the data to be predicted as input to the corresponding trained prediction model to obtain the trajectory prediction result.
2. The method of claim 1, wherein, The preprocessing of each set of collected trajectory data is as follows: Remove invalid values from the speed, position, and heading data, and filter out noise from the speed, position, and heading data.
3. The method of claim 2, wherein, The reconstruction of the position data in each group of preprocessed trajectory data specifically involves: Taking the position data in the m'th group of trajectory data as an example: Step 11: Set the distance threshold; Step 1 and Step 2: Calculate the distance between the location data corresponding to two adjacent time points; If the distance between the location data corresponding to two adjacent time points is greater than the set distance threshold, the two adjacent time points are marked as discontinuities. If the distance between the location data corresponding to two adjacent time points is less than or equal to the set distance threshold, no processing is required; After judging the position data corresponding to every two adjacent time points in the m'th group of trajectory data, all discontinuities in the position data of the m'th group of trajectory are obtained; Step 13: Obtain the number of data points between every two adjacent discontinuities, and compare the obtained number of data points with the trajectory noise threshold. If the number of data points between two adjacent discontinuities is greater than the trajectory noise threshold, then the trajectory data corresponding to the two adjacent discontinuities and the data points between the two adjacent discontinuities are saved as a new set of trajectory data. If the number of data points between two adjacent discontinuities is less than or equal to the trajectory noise threshold, then the trajectory data corresponding to the two adjacent discontinuities and the data points between the two adjacent discontinuities are considered as noise and are removed. 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 processing the position data in each set of trajectory data respectively, the updated trajectory data and the saved new trajectory data are taken as the processed sets of trajectory data.
4. The aircraft trajectory prediction method based on multi-attention collaborative network in a multi-dimensional variable scene according to claim 3, characterized in that, The clustering of the processed sets of trajectory data is based on the position data in the sets of trajectory data, and the clustering process is: Step two one, calculating a distance matrix based on the position data in the processed sets of trajectory data : wherein the trajectory set is a set composed of processed trajectory data groups, is a trajectory set , is the , is the number of trajectory data groups in the set , is the similarity between the first trajectory data group and the trajectory data group in the set . Step two, initialize the number of cluster centers ; Step two three, grouping the trajectories in the set of trajectories according to the distance matrix classes; Step two four, calculate the clustering error according to the clustering result If the clustering error converges, the final clustering result is obtained, and if the clustering error does not converge, let Step two three is executed again.
5. The method of claim 4, wherein, The set The similarity between the first group of trajectory data and the second group of trajectory data is calculated by the following method: The similarity between the first group of trajectory data and the second group of trajectory data is calculated by the following method: wherein is an intermediate variable, is a position point in the 1st group of trajectory data, is a position point in the 1st group of trajectory data; 。 6. The method of claim 5, wherein, The clustering error The calculation process is as follows: in, Indicates the first One cluster, represent Location data in any set of trajectory data, Belongs to The centroid of the location data in all group trajectory data. It represents the total number of categories.
7. The method of claim 6, wherein, The working process of the T-net branch is: The working process of the T-net branch is: Step 1, the hidden state and the target variable sequence as input to the first knowledge-augmented LSTM unit in the encoder module, i.e. ; then wherein, is the hidden state of the -1th knowledge-augmented LSTM unit output in the encoder module, is the candidate state of the -1th knowledge-augmented LSTM unit in the encoder module; Based on the hidden state and Generate the first Hidden states of a knowledge-enhanced LSTM unit: wherein, denotes the transpose of a matrix, is a weight parameter, is an auxiliary context vector, is an importance score of is a weight of wherein and are weight parameters, is the hidden state output of the th knowledge-augmented LSTM unit in the encoder module; Step 2, the decoder module is trained to generate the output sequence As input to each knowledge-augmented LSTM unit in the decoder module, for the th knowledge-augmented LSTM unit in the decoder module, a context vector is generated based on the time attention mechanism, and the hidden state of the th knowledge-augmented LSTM unit in the decoder module : wherein, and are weight parameters, is an element-wise multiplication, is an importance score of is an attention weight of According to The context vector is generated by weighted summation : Step 3, place the first The hidden state of the knowledge-enhanced LSTM unit is passed through a multilayer perceptron, and the output of the multilayer perceptron is the first... The prediction results of the knowledge-enhanced LSTM unit will be used to... The prediction results of each knowledge-enhanced LSTM unit and the context vector Perform cascading operations; Cascade operation result, hidden state and hidden state As the first The input of a knowledge-enhanced LSTM unit, i.e. ; wherein, is a concatenation operation, is a prediction of the kth knowledge-augmented LSTM unit in the decoder module, is a hidden state of the kth knowledge-augmented LSTM unit in the decoder module, is a prediction of the kth knowledge-augmented LSTM unit in the decoder module, is a hidden state of the kth knowledge-augmented LSTM unit in the decoder module, The hidden state is passed through a multi-layer perceptron to output a prediction of the knowledge-augmented LSTM unit.
8. The method of claim 7, wherein, The hidden state Through the multi-layer perceptron, output the prediction result of the knowledge-enhanced LSTM unit, specifically: wherein, is a weight parameter, is the prediction result of the th knowledge-augmented LSTM unit.