High-speed aircraft trajectory prediction method, device and equipment based on parameter estimation
Through the combination of Kalman filtering and sparse correlation structure model, the problems of nonlinear parameter characterization and noise influence in high-speed aircraft trajectory prediction are solved, and the accurate trajectory prediction and generalization improvement of high-speed aircraft are achieved.
Patent Information
- Application Number
- CN202411145145.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-20
- Publication Date
- 2024-11-22
- Estimated Expiration
- 2044-08-20
AI Technical Summary
The prior art has problems in the prediction of trajectory of high-speed aircraft, such as difficulty in characterizing nonlinear parameters, dependence on a large number of trajectory data, great influence on noise and disturbance, and weak generalization and applicability.
Kalman filtering is used to process historical trajectory data, combine random recursive neural networks and sparse correlation structure models for parameter estimation, and reconstruct future trajectories by using cyclic multi-step prediction method to enhance their resistance to noise.
Accurate trajectory prediction of high-speed aircraft is realized, prediction accuracy and model generalization under a small amount of trajectory data are improved, and the impact of noise and uncertain factors is effectively dealt with.
Smart Images

Figure CN118656711B_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of aircraft technology, and in particular to a method, device and equipment for predicting high-speed aircraft trajectories based on parameter estimation. Background Art
[0002] Aircraft trajectory prediction is of great significance for improving flight safety, optimizing flight plans, improving airspace management efficiency, supporting air defense and anti-missile systems, and enhancing search and rescue and emergency response capabilities.
[0003] Existing aircraft trajectory prediction technologies are mainly divided into two categories: one is the trajectory prediction for low-altitude and low-speed targets such as aircraft and drones. Most of the literature uses methods such as database matching, deep learning, and machine learning. The other is the trajectory prediction for high-altitude and high-speed targets such as ballistic missiles. Existing methods use analytical methods, numerical integration, and regression analysis to solve. Although the existing technology has made certain progress in the trajectory prediction of low-altitude and low-speed targets, these methods have limitations when applied to high-speed vehicles (HGVs). Specifically, there are the following limitations: when using linear fitting algorithms such as the least squares method to predict parameters, it is difficult to accurately characterize and predict parameters with nonlinear laws, and the use of deep learning technology to predict requires a large amount of trajectory data support, which is not supported by real scenarios. In addition, HGV trajectory data will be affected by uncertain factors such as noise, disturbance, and environment. When used directly for prediction, the model prediction accuracy is low. At the same time, most methods are based on the premise that the control quantity or parameter has a certain law, and the universality and applicability of the methods are weak. Summary of the invention
[0004] Based on this, it is necessary to provide a high-speed aircraft trajectory prediction method, device and equipment based on parameter estimation that can accurately predict the trajectory of high-speed moving aircraft in order to solve the above technical problems.
[0005] A high-speed aircraft trajectory prediction method based on parameter estimation, the method comprising:
[0006] Acquire historical trajectory data of the target aircraft, and process the historical trajectory data using Kalman filtering to obtain historical observation data;
[0007] Using a trained parameter prediction model to perform parameter estimation based on the historical observation data, to obtain trajectory prediction parameters including lift acceleration, drag acceleration and roll angle, wherein in the trained parameter prediction model, the historical observation data is used to obtain corresponding latent variable data through a random recursive neural network, the latent variable data is connected to the historical observation data to obtain new data, and a sparse association structure model is used to perform parameter estimation based on the new data to obtain the trajectory prediction parameters;
[0008] According to the trajectory prediction parameters and the kinematic equation of the high-speed aircraft, a cyclic multi-step prediction method is adopted to obtain the future predicted trajectory of the target aircraft through integral extrapolation reconstruction.
[0009] In one embodiment, the historical observation data includes time series data corresponding to a plurality of state variables.
[0010] In one embodiment, connecting the latent variable data to the historical observation data includes:
[0011] The random recurrent neural network generates hidden variable data corresponding to each state variable according to the time series data of each state variable;
[0012] The latent variable data is added to the time series data of the corresponding state variable to form the newly added data.
[0013] In one embodiment, when the sparse association structure model is used to perform parameter estimation based on the newly added data:
[0014] Divide the time series data of each state variable into multiple subsequences with increasing lengths;
[0015] Adopting an independent long-short memory network layer and an attention mechanism based on the state variables, adaptive sequence representation is performed according to the plurality of subsequences to obtain adaptive representations corresponding to the state variables;
[0016] Reconstructing the sparse association structure between the state variables according to the adaptive representations based on the attention mechanism between the state variables;
[0017] Obtaining a final representation structure according to the sparse association structure obtained after reconstruction and the adaptive representation of each of the state variables;
[0018] A predictor is used to perform parameter estimation according to the final representation structure to obtain the trajectory prediction parameters.
[0019] In one embodiment, when training the parameter prediction model:
[0020] Simulate according to various types of aircraft to obtain multiple sets of simulation observation data;
[0021] Calculate the lift acceleration, drag acceleration and roll angle corresponding to each group of simulation observation data and use them as true value labels;
[0022] Based on each group of simulation observation data and the corresponding true value labels, a training data set for training the parameter prediction model is constructed.
[0023] In one embodiment, when training the parameter prediction model, the parameter prediction model is trained using mean square error and KL divergence loss until convergence to obtain the trained parameter prediction model, wherein L2 regularization is added to the mean square error.
[0024] The present application also provides a high-speed aircraft trajectory prediction device based on parameter estimation, the device comprising:
[0025] A historical observation data acquisition module is used to acquire historical trajectory data of the target aircraft, and process the historical trajectory data using Kalman filtering to obtain historical observation data;
[0026] A trajectory parameter prediction module, used to perform parameter estimation based on the historical observation data using a trained parameter prediction model to obtain trajectory prediction parameters including lift acceleration, drag acceleration and roll angle, wherein in the trained parameter prediction model, the historical observation data is used to obtain corresponding latent variable data through a random recursive neural network, the latent variable data is connected to the historical observation data to obtain new data, and a sparse association structure model is used to perform parameter estimation based on the new data to obtain the trajectory prediction parameters;
[0027] The future trajectory prediction module is used to obtain the future predicted trajectory of the target aircraft by integrating and extrapolating the reconstruction based on the trajectory prediction parameters and the kinematic equation of the high-speed aircraft using a cyclic multi-step prediction method.
[0028] A computer device comprises a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the following steps are implemented:
[0029] Acquire historical trajectory data of the target aircraft, and process the historical trajectory data using Kalman filtering to obtain historical observation data;
[0030] Using a trained parameter prediction model to perform parameter estimation based on the historical observation data, to obtain trajectory prediction parameters including lift acceleration, drag acceleration and roll angle, wherein in the trained parameter prediction model, the historical observation data is used to obtain corresponding latent variable data through a random recursive neural network, the latent variable data is connected to the historical observation data to obtain new data, and a sparse association structure model is used to perform parameter estimation based on the new data to obtain the trajectory prediction parameters;
[0031] According to the trajectory prediction parameters and the kinematic equation of the high-speed aircraft, a cyclic multi-step prediction method is adopted to obtain the future predicted trajectory of the target aircraft through integral extrapolation reconstruction.
[0032] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the following steps:
[0033] Acquire historical trajectory data of the target aircraft, and process the historical trajectory data using Kalman filtering to obtain historical observation data;
[0034] Using a trained parameter prediction model to perform parameter estimation based on the historical observation data, to obtain trajectory prediction parameters including lift acceleration, drag acceleration and roll angle, wherein in the trained parameter prediction model, the historical observation data is used to obtain corresponding latent variable data through a random recursive neural network, the latent variable data is connected to the historical observation data to obtain new data, and a sparse association structure model is used to perform parameter estimation based on the new data to obtain the trajectory prediction parameters;
[0035] According to the trajectory prediction parameters and the kinematic equation of the high-speed aircraft, a cyclic multi-step prediction method is adopted to obtain the future predicted trajectory of the target aircraft through integral extrapolation reconstruction.
[0036] The above-mentioned high-speed aircraft trajectory prediction method, device and equipment based on parameter estimation uses Kalman filtering to process the historical trajectory data of the target aircraft to obtain historical observation data, and uses the trained parameter prediction model to perform parameter estimation based on the historical observation data to obtain trajectory prediction parameters including lift acceleration, drag acceleration and roll angle, wherein in the trained parameter prediction model, the historical observation data is used to obtain corresponding hidden variable data through a random recursive neural network, the hidden variable data is connected to the historical observation data to obtain new data, and the sparse association structure model is used to perform parameter estimation based on the new data to obtain trajectory prediction parameters, and according to the trajectory prediction parameters and the kinematic equation of the high-speed aircraft, the cyclic multi-step prediction method is used to reconstruct the future predicted trajectory of the target aircraft through integral extrapolation. This method can be used to accurately predict the trajectory of high-speed aircraft. BRIEF DESCRIPTION OF THE DRAWINGS
[0037] Figure 1 A schematic diagram of a flow chart of a high-speed aircraft trajectory prediction method based on parameter estimation in one embodiment;
[0038] Figure 2 FIG. 1 is a schematic diagram of a longitudinal maneuver of a high-speed aircraft in one embodiment, wherein: Figure 2 (a) is a schematic diagram of longitudinal maneuvers during balanced gliding. Figure 2 (b) Schematic diagram of longitudinal maneuver during jump gliding;
[0039] Figure 3 A schematic diagram of a SASM algorithm model in one embodiment;
[0040] Figure 4 Schematic diagram of the structure of the SRNN algorithm model in one embodiment, wherein: Figure 4 (a) shows the generation model in SRNN. Figure 4 (b) shows the variational inference network in SRNN;
[0041] Figure 5 A schematic diagram of the structure of a parameter prediction model in one embodiment;
[0042] Figure 6 A schematic diagram of cyclic multi-step prediction in one embodiment;
[0043] Figure 7 It is a flowchart of a high-speed aircraft trajectory prediction method based on parameter estimation in another embodiment;
[0044] Figure 8 It is a schematic diagram of the error of the roll angle prediction of aircraft model 1 in an experiment using the parameter prediction model in this method;
[0045] Fig. 9 It is a schematic diagram of the error of lift acceleration prediction of aircraft model 1 in an experiment using the parameter prediction model in this method;
[0046] Fig.10 It is a schematic diagram of the error of drag acceleration prediction of aircraft model 1 in an experiment using the parameter prediction model in this method;
[0047] Fig.11 It is a schematic diagram of the error of the roll angle prediction of aircraft model 2 in an experiment using the parameter prediction model in this method;
[0048] Fig.12 It is a schematic diagram of the error of lift acceleration prediction of aircraft model 2 in an experiment using the parameter prediction model in this method;
[0049] Fig.13 It is a schematic diagram of the error of drag acceleration prediction of aircraft model 2 in an experiment using the parameter prediction model in this method;
[0050] Fig.14 is a structural block diagram of a high-speed aircraft trajectory prediction device based on parameter estimation in one embodiment;
[0051] Fig.15 FIG. 4 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION
[0052] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.
[0053] In view of the many defects in the existing technology when predicting the trajectory of high-speed aircraft, for example, when using linear fitting algorithms such as the least squares method to predict parameters, it is difficult to accurately characterize and predict parameters with nonlinear laws, and the use of deep learning technology for prediction requires a large amount of trajectory data support, which is not supported by real scenarios. At the same time, the trajectory data of high-speed aircraft will be affected by uncertainty factors such as noise, disturbance, and environment. When used directly for prediction, the model prediction accuracy is low, and most methods are based on the premise that the control quantity or parameter has a certain law, and the universal applicability of the method is weak. In this application, Figure 1 As shown, a high-speed aircraft trajectory prediction method based on parameter estimation is provided, comprising the following steps:
[0054] Step S100, obtaining historical trajectory data of the target aircraft, and processing the historical trajectory data using Kalman filtering to obtain historical observation data.
[0055] Step S110, using the trained parameter prediction model to perform parameter estimation based on the historical observation data to obtain trajectory prediction parameters including lift acceleration, drag acceleration and roll angle, wherein in the trained parameter prediction model, the historical observation data is used to obtain corresponding latent variable data through a random recursive neural network, the latent variable data is connected to the historical observation data to obtain new data, and the sparse association structure model is used to perform parameter estimation based on the new data to obtain trajectory prediction parameters.
[0056] In step S120, based on the trajectory prediction parameters and the kinematic equation of the high-speed aircraft, a cyclic multi-step prediction method is used to reconstruct the future predicted trajectory of the target aircraft through integral extrapolation.
[0057] In this embodiment, a method for predicting the trajectory of a speeding aircraft based on parameter estimation is provided. The correlation between historical trajectory data and parameters to be predicted is mined based on a trained parameter prediction model to predict future trajectory parameters. Then, trajectory reconstruction is achieved by performing integral extrapolation based on the aircraft motion equation. This method can accurately predict the trajectory of a completely new and unknown aircraft with relatively few known trajectory data samples, and can effectively deal with the influence of uncertain factors such as noise on trajectory prediction.
[0058] In this embodiment, the object of trajectory prediction is a high-speed aircraft, whose speed can reach Mach 15.
[0059] In this embodiment, the trajectory prediction method adopted is designed based on the trajectory prediction method of parameter identification. First, it is necessary to reasonably select the trajectory parameters, that is, to find parameters with stable rules and control information as the parameters to be estimated, and to describe them mathematically. From the perspective of HGV (high-speed vehicle) maneuvering model cognition, appropriate trajectory prediction parameters are selected.
[0060] Specifically, considering the earth as a uniform sphere and ignoring its rotation, the six-degree-of-freedom dynamic equation of a high-speed glider in a half-velocity coordinate system (VTC) can be expressed as follows:
[0061] (1)
[0062] In formula (1), , , , , , Represents the state variables of the aircraft, which respectively represent longitude, latitude, distance from the center of the earth, speed, speed inclination and track yaw angle. Among them, , are control variables, representing the angle of attack and the roll angle, , They represent lift acceleration and drag acceleration respectively, and the calculation formula is as follows:
[0063] (2)
[0064] In formula (2), , represent lift and drag respectively, denote the lift coefficient and the drag coefficient, respectively. , They represent the equivalent cross-sectional area and mass of the aircraft respectively, Represents the density of the atmosphere.
[0065] Furthermore, the maneuvering characteristics of high-speed aircraft are analyzed. The main maneuvering modes of high-speed aircraft include longitudinal maneuvering, lateral maneuvering and the combination of the two. Among them, longitudinal maneuvering is mainly caused by gravity and aerodynamic lift, and mainly includes two forms: jump gliding and balanced gliding, such as Figure 2 The main attack angle of a high-speed aircraft in the longitudinal plane is and ballistic inclination The impact of When , the aircraft is in a balanced gliding state, as shown in formula (3). In fact, jumping gliding is more conducive to the aircraft's penetration. Therefore, in this embodiment, the jumping gliding state of the high-speed aircraft is mainly studied.
[0066] (3)
[0067] Next, the lateral maneuver of the high-speed aircraft is analyzed. Affected by the roll angle, the lift of the aircraft can be decomposed into climbing force and turning force, and the lateral force is mainly affected by the turning force. By controlling the change of the roll angle, different lateral maneuvering modes of the aircraft can be achieved, such as C-shaped maneuver and S-shaped maneuver. When the roll angle is zero, the turning force is also zero, and the aircraft does not maneuver in the lateral direction.
[0068] In this embodiment, when selecting trajectory prediction parameters, the following two aspects need to be considered: (1) whether the prediction parameters contain control information for the aircraft; (2) whether the prediction parameters have a relatively stable change pattern, which is convenient for mathematical description. Through the above research on the maneuvering characteristics of HGV, it can be seen that the aerodynamic lift acceleration and aerodynamic drag acceleration can fully contain the control variables, and also contain unknown information such as the aircraft mass and area, and the aerodynamic acceleration has a stable change pattern, such as Figure 2 As shown in the formula (1), it can be seen that the change of the roll angle will directly affect the ballistic inclination angle and ballistic deviation angle of the aircraft, and is the main determining factor of the lateral maneuvering mode of the aircraft. Therefore, in this method, the lift acceleration, drag acceleration and roll angle are selected as the prediction parameters of the high-speed aircraft trajectory. Therefore, in step S110, the lift acceleration, drag acceleration and roll angle are predicted according to the historical trajectory data of the aircraft.
[0069] In step S100, the historical trajectory data obtained is the echo data of the target aircraft to be predicted for trajectory observed by radar at a previous time, and the data includes time series data corresponding to multiple state variables. Among them, the state variables refer to the longitude, latitude, distance from the center of the earth, speed, speed inclination and speed azimuth in formula (1). After obtaining the historical trajectory data, Kalman filtering is performed on it to obtain historical observation data, and parameter prediction is performed based on the historical observation data.
[0070] Furthermore, considering that the trajectory sequence is actually a time series, the time domain adaptive method is introduced into the prediction of aircraft trajectory parameters. Traditional time series processing methods (such as RNN, LSTM, etc.) mostly ignore the correlation structure between variables and are prone to overfitting problems.
[0071] Therefore, in this embodiment, a deep learning method based on causal inference, the Sparse Associative Structure Model (SASM), is used to mine the causal relationship between data, remove the pseudo correlation between variables, and further improve the generalization and accuracy of the model. At the same time, in order to reduce the impact of uncertainty factors such as noise on the prediction accuracy of the model, the Stochastic Recurrent Neural Networks (SRNN) algorithm is also introduced to use its random layer modeling data uncertainty problems to further improve the prediction accuracy.
[0072] Specifically, the core idea of the Sparse Association Structure Model (SASM) is to relax the causal relationship to a sparse association structure based on the causal sparsity criterion, and then use the sequence representation and association capture capabilities of deep learning to perform approximate causal discovery by reconstructing the sparse association structure. The model is mainly divided into two parts: adaptive sequence representation and sparse association structure discovery. Figure 3 shown.
[0073] Assume that the input trajectory data is , among which The time series data corresponding to the variables is First, we need to determine the candidate segment set, that is, a subsequence of the original time sequence, expressed as In the sparse association structure model, each state variable is independently divided into a corresponding LSTM layer (i.e., long short-term memory network layer), which makes the model training process parallel and can take into account the information at all times at the same time.
[0074] Specifically, given The time step feature segments, we can get the characterization of the candidate segments: ,in Implemented by the LSTM layer, Indicates LSTM parameters. It is worth noting that any univariate time series The segments all share the same LSTM network, and finally we can get the segment representation set as shown below: Then, the attention mechanism is used to focus on the precise segment representation among all candidate segment representations. In form, the calculation The weights of the segments are as follows:
[0075] (4)
[0076] (5)
[0077] In formula (4), are the trainable projection parameters of different fully connected layers (FC), which are composed of linear layers and activation functions superimposed on each other. is the scaling factor. In order to obtain sparse weights that clearly represent a specific segment, the sparsemax activation function is selected to calculate the weights. Compared with the softmax activation function, sparsemax can obtain more sparse weights, so that the final adaptive representation pays more attention to a specific segment. Finally, we can get Adaptive representation of variables :
[0078] (6)
[0079] In formula (6), Expressed as length The weight of the candidate temporal segment representation, are the trainable projection parameters of FC.
[0080] Furthermore, considering that it is difficult to learn the true causal relationship between parallel connections, it is necessary to relax the causal relationship to a sparse correlation structure. After obtaining the adaptive representation of each variable, these representations can be used to reconstruct the sparse correlation structure between variables. and variables The degree of correlation between The method is shown below:
[0081] (7)
[0082] Use sparsemax to sparse these correlation levels:
[0083] (8)
[0084] In formula (8), For variables With variables The degree of correlation For variables A collection of correlations with other variables. It can be regarded as the correlation between the latitude feature and the speed feature in the target data set. It can be regarded as a collection of the degree of correlation between latitude features and other features such as speed. Therefore, the sparse correlation between different variables can be obtained through the above method, and the collection of the degree of correlation between different variables constitutes the correlation structure between variables, that is, the sparse correlation structure.
[0085] Different variables have different effects on the prediction results. In a sparse structure, the more correlated a variable is with other variables, the more important it is. Therefore, the sparse correlation representation of different variables can be used to control the influence of the adaptive representation of the variable on the final representation, thereby affecting the prediction results. The structural representation of the association The calculation of is shown as follows: For a linear layer:
[0086] (9)
[0087] Then, get the Controlled representation of variables :
[0088] (10)
[0089] In formula (10), is the sigmoid function, which maps the input value to , is the Hadamard product. Controlled representation Is subject to variable Corresponding adaptive representation The structural representation of the association Assuming that the velocity correlation is interpretable in the aircraft trajectory dataset, the sparse correlation structure of latitude Via the network Get the correlation structure representation of latitude , after the sigmoid function Map to range, used to control the adaptive representation of latitude degree of impact.
[0090] Then, the final representation structure of the data is obtained , which is connected by controlled representations of different variables.
[0091] Finally, the final representation structure is input into the predictor The final prediction result is obtained:
[0092]
[0093] When training the sparse correlation structure model, the optimized mean square error (MSE) is used as the optimization objective function:
[0094]
[0095] In formula (12), the second term (the term after the plus sign) is the L2 regularization factor, represents the regularization parameter, Represents the weight parameter of the model. L2 regularization helps solve the overfitting problem of the model.
[0096] Furthermore, based on the sparse correlation structure model, the uncertainty factors in the target trajectory are fully considered, and it is proposed to use the random layer of the stochastic recurrent neural network (SRNN) to model the uncertainty problem in the historical observation data in order to improve the accuracy of the model.
[0097] like Figure 4 As shown in Figure 1, SRNN (Stochastic Recurrent Neural Network) is a deep learning model that combines determinism and randomness to process sequence data. The SRNN model is mainly composed of a generative model (see Figure 4 (a)) and the variational inference network (see Figure 4 (b)) consists of two parts.
[0098] Figure 4 (a) shows the SRNN-based sequence Generative Model , It combines recurrent neural networks (RNNs) and state space models (SSMs). It should be emphasized that the random state Depends on deterministic state , and vice versa, which ensures that the deterministic state is separated from the random state and can resist randomness. Therefore, the joint density probability distribution of the generative model can be rewritten as follows:
[0099] (13)
[0100] In formula (13), Represents a random state given the previous time step and deterministic state Under the condition of The prior transfer distribution of is a Gaussian distribution with a covariance diagonal matrix. The mean and log variance of the distribution are parameterized by neural networks NN1 and NN2, that is:
[0101] (14)
[0102] Furthermore, Figure 4 (b) Describes the variational inference network of SRNN The network can obtain hidden variables and The approximate posterior distribution of is as follows:
[0103] (15)
[0104] Due to the deterministic state Depends only on and , the latter term in formula (15) can be expressed as , can be estimated based on past outputs and deterministic states In order to express future information, SRNN introduces a reverse recursive state, and its update formula is The first term of formula (15) can be further expressed as:
[0105] (16)
[0106] In formula (16), So as well as is a Gaussian distribution with parameters.
[0107] In this embodiment, in order to improve the accuracy of the model's on-orbit prediction, LSTM units are used instead of RNN units in SRNN.
[0108] When training SRNN, mean square error and KL divergence loss are used to measure the difference between the model output and the real data. The KL divergence calculation formula is as follows:
[0109] (17)
[0110] In formula (17), Represents the variance and mean of the output distribution and the true distribution. At the same time, the Adam algorithm is used as the optimizer to update the model parameters.
[0111] In step S110, the parameter prediction model used to estimate the trajectory prediction parameters is a combination of the above two models. On the basis of SASM, the data is first input into the SRNN and processed, and then the latent variables obtained by the SRNN are connected with the original variables and input into the SASM model together, as shown in FIG. Figure 5 shown.
[0112] Furthermore, connecting the latent variable data to the historical observation data includes: a random recurrent neural network generates latent variable data corresponding to each state variable according to the time series data of each state variable, and adding the latent variable data to the time series data of the corresponding state variable to form newly added data.
[0113] Specifically, when the sparse association structure model is used to perform parameter estimation based on the newly added data: the time series data of each state variable is divided into multiple subsequences with increasing lengths, and an independent length-short memory network layer and an attention mechanism within the state variable are used to perform adaptive sequence representation based on the multiple subsequences to obtain adaptive representations corresponding to the state variables, and then the sparse association structure between the state variables is reconstructed according to each adaptive representation based on the attention mechanism between the state variables, and the final representation structure is obtained based on the sparse association structure obtained after reconstruction and the adaptive representation of each state variable. Finally, the predictor is used to perform parameter estimation based on the final representation structure to obtain trajectory prediction parameters.
[0114] In this embodiment, SRNN is used to extract latent variable data of each state variable in the historical observation data, and the diversity of the original state variables can be increased, so that more data is input into the SASM model, and the SASM model has a better prediction effect.
[0115] In this embodiment, when training the parameter prediction model, taking into account the diverse categories of target aircraft in real application scenarios, in order to improve the generalization of the parameter prediction model, when constructing the training data set: first, simulation is performed according to various aircraft categories to obtain multiple groups of simulation observation data, and calculations are performed based on each group of simulation observation data to obtain the lift acceleration, drag acceleration and roll angle corresponding to each group of simulation observation data, and these are used as true value labels. Based on each group of simulation observation data and the corresponding true value labels, a training data set for training the parameter prediction model is constructed.
[0116] In this embodiment, when the trained parameter prediction model is put into practical application, trajectory prediction can be performed for aircraft of unknown models, where the unknown model refers to a model that does not appear in the training data set. The trained parameter prediction model can migrate the model trained based on existing data to unknown data, and accurately predict the parameters of the unknown model aircraft.
[0117] In this embodiment, when training the parameter prediction model, the mean square error formula (12) and the KL divergence loss formula (17) are used to train the parameter prediction model until convergence, thereby obtaining a trained parameter prediction model.
[0118] In step S120, after completing the prediction of the trajectory parameters, it is necessary to reconstruct the trajectory through integral extrapolation based on the six-degree-of-freedom motion equation formula (1) and the current observation data of the high-speed aircraft to obtain the predicted trajectory.
[0119] In this embodiment, a cyclic multi-step prediction method is used to predict the future trajectory based on the predicted parameters, such as Figure 6As shown in the figure, the trajectory data at the moment is input in the tth step, and after the SASM-SRNN model, i.e. the parameter prediction model, the prediction output at the moment t is obtained. The prediction output is input into the model together with the input data at the moment t+1 for prediction, and the prediction at the moment t+1 is obtained. This cycle is repeated to realize the trajectory prediction for multiple time steps in the future.
[0120] In this embodiment, the overall process of the high-speed aircraft trajectory prediction method based on parameter estimation proposed in this paper is as follows: Figure 7 As shown, it includes target aircraft trajectory tracking and data denoising, trajectory parameter selection, parameter prediction and trajectory reconstruction.
[0121] In order to prove the effectiveness of the parameter prediction model in this method, a comparative experiment was conducted with the benchmark algorithm LSTM. The results of the comparative experiment are shown in Table 1. Figure 8-Figure 13 Schematic diagram of the prediction error of lift acceleration, drag acceleration and roll angle for two different types of aircraft.
[0122] Table 1
[0123]
[0124] In the experiment, the trajectory data generated by two common models of aircraft were tested. Each set of training data includes 250 trajectories, and the length of each trajectory is between 700 and 1000 seconds. The mean square error (MSE) and mean absolute error (MAE) of the test samples are used as evaluation indicators, as shown in formulas (18) and (19).
[0125] (18)
[0126] (19)
[0127] In formula (18) and formula (19), represent the true value and the predicted value respectively.
[0128] Furthermore, it can be seen from Table 1 that the mean square error and absolute error of the proposed method in predicting the three parameters of the two types of aircraft are smaller than those of the baseline algorithm. In particular, when predicting the roll angle parameter, the prediction MSE of the proposed method has a maximum error reduction rate of 90.98% compared with the LSTM baseline method, and the MAE reduction rate has reached 76.98%.
[0129] exist Figure 8-Figure 13In the figure, the ordinate represents the true value of the predicted parameter. It can be seen that the error band of the proposed method is smaller than that of the baseline algorithm. Although the proposed method is not significantly better than the baseline algorithm in the prediction of some interval values of aerodynamic lift acceleration and drag acceleration, the error band of the proposed method is significantly smaller than that of the baseline algorithm in the prediction of the roll angle, which proves the effectiveness of the proposed method.
[0130] In the above-mentioned high-speed aircraft trajectory prediction method based on parameter estimation, in view of the problem that the trajectory prediction method based on deep learning in the prior art relies on a large amount of training data, the problem that the existing method for trajectory prediction of unknown models has poor prediction accuracy, and the problem that the commonly used Kalman filter method has limited data processing effect when dealing with the uncertainty of trajectory data, a domain adaptive method in the time domain is proposed for the first time, and the sparse association structure model (SASM) is applied to the trajectory parameter prediction of HGV. First, in the adaptive sequence representation in SASM, each univariate trajectory time series data is divided into multiple candidate segments, which are encoded by a separate LSTM network to obtain a segment representation set of the univariate trajectory time series data. And under the sparsity criterion, the sparse weights of each candidate segment in the segment representation set are calculated with the help of the attention mechanism to obtain the adaptive representation of each segment. The above steps can obtain more fine-grained time segments and can solve the noise of offset caused by the inconsistent start time of different variables affecting other variables. Then, the adaptive sequence representation of each univariate sequence is used to calculate the correlation degree between different variables based on the self-attention mechanism. The set of correlation degrees between different variables constitutes a sparse association structure between variables. The sparse association structure of the variable is then processed by a linear layer and multiplied with its adaptive representation to obtain its final association representation structure representation. Finally, the representation structure is input into the classifier to obtain the final prediction result. This method can mine the association between the feature observations of known data and the trajectory parameters to be predicted based on the existing less trajectory data, and then apply it to the trajectory data of new unknown aircraft models. Through the migration of association relationships, the label quantity of the unknown aircraft can be estimated from the observations.
[0131] At the same time, in order to consider the uncertainty problem in trajectory data, a random recurrent neural network (SRNN) is proposed to model the uncertainty problem in the data. First, the deterministic layer and the uncertainty layer of the model are constructed using the long short-term memory network unit (LSTM) network and the fully connected layer, respectively. Then, based on the variational autoencoder principle, the variational inference network of SRNN is constructed to calculate the posterior approximate distribution of the hidden variables. Finally, the mean square error (RSE) and KL divergence are used as the loss function to calculate the gradient, and Adam is used as the optimizer to update the model parameters.
[0132] The initial trajectory data is first input into the SRNN model, and the trained latent variables are concatenated with the original data along the feature dimension and then input into the SASM model for subsequent prediction. The introduction of SRNN helps to express uncertainty problems with mathematical descriptions, enriching the input feature information, thereby constructing a richer sparse correlation structure containing latent variables and observed variables, and using it to further calculate the final representation with richer information, which is used for prediction results to improve the prediction performance of the model.
[0133] It should be understood that although Figure 1 The steps in the flowchart are shown in sequence as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. Moreover, Figure 1 At least part of the steps may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least part of the sub-steps or stages of other steps.
[0134] In one embodiment, Fig.14 As shown, a high-speed aircraft trajectory prediction device based on parameter estimation is provided, comprising: a historical observation data acquisition module 200, a trajectory parameter prediction module 210 and a future trajectory prediction module 220, wherein:
[0135] The historical observation data acquisition module 200 is used to acquire the historical trajectory data of the target aircraft, and process the historical trajectory data using Kalman filtering to obtain historical observation data;
[0136] A trajectory parameter prediction module 210 is used to perform parameter estimation based on the historical observation data using a trained parameter prediction model to obtain trajectory prediction parameters including lift acceleration, drag acceleration and roll angle, wherein in the trained parameter prediction model, the historical observation data is used to obtain corresponding latent variable data through a random recursive neural network, the latent variable data is connected to the historical observation data to obtain new data, and a sparse association structure model is used to perform parameter estimation based on the new data to obtain the trajectory prediction parameters;
[0137] The future trajectory prediction module 220 is used to obtain the future predicted trajectory of the target aircraft by integral extrapolation reconstruction using a cyclic multi-step prediction method according to the trajectory prediction parameters and the kinematic equation of the high-speed aircraft.
[0138] For the specific definition of the high-speed aircraft trajectory prediction device based on parameter estimation, please refer to the definition of the high-speed aircraft trajectory prediction method based on parameter estimation in the above text, which will not be repeated here. Each module in the above-mentioned high-speed aircraft trajectory prediction device based on parameter estimation can be implemented in whole or in part by software, hardware and a combination thereof. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.
[0139] In one embodiment, a computer device is provided. The computer device may be a terminal, and its internal structure diagram may be as follows: Fig.15 As shown. The computer device includes a processor, a memory, a network interface, a display screen and an input device connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a high-speed aircraft trajectory prediction method based on parameter estimation is implemented. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device can be a touch layer covering the display screen, or a button, trackball or touchpad set on the computer device housing, or an external keyboard, touchpad or mouse, etc.
[0140] Those skilled in the art will understand that Fig.15 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.
[0141] In one embodiment, a computer device is provided, including a memory and a processor, wherein a computer program is stored in the memory, and when the processor executes the computer program, the following steps are implemented:
[0142] Acquire historical trajectory data of the target aircraft, and process the historical trajectory data using Kalman filtering to obtain historical observation data;
[0143] Using a trained parameter prediction model to perform parameter estimation based on the historical observation data, to obtain trajectory prediction parameters including lift acceleration, drag acceleration and roll angle, wherein in the trained parameter prediction model, the historical observation data is used to obtain corresponding latent variable data through a random recursive neural network, the latent variable data is connected to the historical observation data to obtain new data, and a sparse association structure model is used to perform parameter estimation based on the new data to obtain the trajectory prediction parameters;
[0144] According to the trajectory prediction parameters and the kinematic equation of the high-speed aircraft, a cyclic multi-step prediction method is adopted to obtain the future predicted trajectory of the target aircraft through integral extrapolation reconstruction.
[0145] In one embodiment, a computer readable storage medium is provided, on which a computer program is stored, and when the computer program is executed by a processor, the following steps are implemented:
[0146] Acquire historical trajectory data of the target aircraft, and process the historical trajectory data using Kalman filtering to obtain historical observation data;
[0147] Using a trained parameter prediction model to perform parameter estimation based on the historical observation data, to obtain trajectory prediction parameters including lift acceleration, drag acceleration and roll angle, wherein in the trained parameter prediction model, the historical observation data is used to obtain corresponding latent variable data through a random recursive neural network, the latent variable data is connected to the historical observation data to obtain new data, and a sparse association structure model is used to perform parameter estimation based on the new data to obtain the trajectory prediction parameters;
[0148] According to the trajectory prediction parameters and the kinematic equation of the high-speed aircraft, a cyclic multi-step prediction method is adopted to obtain the future predicted trajectory of the target aircraft through integral extrapolation reconstruction.
[0149] Those of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be implemented by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.
[0150] The technical features of the above embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.
[0151] The above-mentioned embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the attached claims.
Claims
1. A high-speed aircraft trajectory prediction method based on parameter estimation, characterized in that: The method comprises: Acquire historical trajectory data of the target aircraft, and process the historical trajectory data using Kalman filtering to obtain historical observation data including time series data corresponding to a plurality of state variables; Parameters are estimated based on the historical observation data using a trained parameter prediction model to obtain trajectory prediction parameters including lift acceleration, drag acceleration and roll angle, wherein in the trained parameter prediction model, the historical observation data is used to obtain corresponding latent variable data through a random recurrent neural network, the latent variable data is connected to the historical observation data to obtain new data, and parameters are estimated based on the new data using a sparse association structure model to obtain the trajectory prediction parameters, wherein the connection of the latent variable data to the historical observation data includes: the random recurrent neural network generates latent variable data corresponding to each state variable based on the time series data of each state variable, The latent variable data is added to the time series data of the corresponding state variable to form the newly added data. When the sparse association structure model is used to perform parameter estimation based on the newly added data: the time series data of each state variable is divided into a plurality of subsequences with increasing lengths in sequence, an independent length-short memory network layer and an attention mechanism within the state variable are used to perform adaptive sequence representation based on the plurality of subsequences to obtain adaptive representations corresponding to the state variables, the sparse association structure between the state variables is reconstructed based on the adaptive representations based on the attention mechanism between the state variables, a final representation structure is obtained based on the sparse association structure obtained after reconstruction and the adaptive representations of the state variables, and a predictor is used to perform parameter estimation based on the final representation structure to obtain the trajectory prediction parameters; According to the trajectory prediction parameters and the kinematic equation of the high-speed aircraft, a cyclic multi-step prediction method is adopted to obtain the future predicted trajectory of the target aircraft through integral extrapolation reconstruction.
2. The high-speed aircraft trajectory prediction method according to claim 1, characterized in that: When training the parameter prediction model: Simulate according to various types of aircraft to obtain multiple sets of simulation observation data; Calculate the lift acceleration, drag acceleration and roll angle corresponding to each group of simulation observation data and use them as true value labels; Based on each group of simulation observation data and the corresponding true value labels, a training data set for training the parameter prediction model is constructed.
3. The high-speed aircraft trajectory prediction method according to claim 2, characterized in that: When training the parameter prediction model, the parameter prediction model is trained using mean square error and KL divergence loss until convergence to obtain the trained parameter prediction model, wherein L2 regularization is added to the mean square error.
4. A high-speed aircraft trajectory prediction device based on parameter estimation, characterized in that: The device comprises: A historical observation data acquisition module is used to acquire historical trajectory data of the target aircraft, and process the historical trajectory data using Kalman filtering to obtain historical observation data including time series data corresponding to a plurality of state variables; The trajectory parameter prediction module is used to use the trained parameter prediction model to perform parameter estimation based on the historical observation data to obtain trajectory prediction parameters including lift acceleration, drag acceleration and roll angle, wherein in the trained parameter prediction model, the historical observation data is used to obtain corresponding latent variable data through a random recurrent neural network, the latent variable data is connected to the historical observation data to obtain new data, and a sparse association structure model is used to perform parameter estimation based on the new data to obtain the trajectory prediction parameters, wherein the connection of the latent variable data to the historical observation data includes: the random recurrent neural network generates latent variable data corresponding to each state variable according to the time series data of each state variable, and the latent variable data is connected to the historical observation data to obtain new data. Adding to the time series data of the corresponding state variables to form the newly added data, when the sparse association structure model is used to perform parameter estimation based on the newly added data: the time series data of each state variable is divided into a plurality of subsequences with increasing lengths, an independent length-short memory network layer and an attention mechanism within the state variables are used to perform adaptive sequence representation according to the plurality of subsequences to obtain adaptive representations corresponding to the state variables, the sparse association structure between the state variables is reconstructed according to the adaptive representations based on the attention mechanism between the state variables, a final representation structure is obtained according to the sparse association structure obtained after reconstruction and the adaptive representations of the state variables, and a predictor is used to perform parameter estimation according to the final representation structure to obtain the trajectory prediction parameters; The future trajectory prediction module is used to obtain the future predicted trajectory of the target aircraft by integrating and extrapolating the reconstruction based on the trajectory prediction parameters and the kinematic equation of the high-speed aircraft using a cyclic multi-step prediction method.
5. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 3 are implemented.
