Track prediction intelligent optimization system and method

Abstract

The invention discloses an intelligent optimization system and method for track prediction. The method comprises the steps that 1, TLE data are preprocessed, and a data set is obtained; 2, training a machine learning model; 3, propagating the current state through aircraft dynamics by adopting an estimation algorithm, and then correcting the propagation state by utilizing a pseudo measurement method; 4, evaluating and optimizing the model; and 5, predicting and classifying. Machine learning prediction is used as UKF-based pseudo measurement prediction, a current state is propagated through an aircraft dynamics model, then the propagation state is corrected by using pseudo measurement data, and then an HMMS is used for predicting a pseudo measurement value driven by data, so that accurate estimation of a future state of an aircraft can be realized by effectively combining the prediction capability of machine learning. The method not only considers the similarity of the trajectory in the spatial dimension, but also can effectively capture the dynamic change in the time dimension, thereby more accurately predicting the flight trajectory.

Description

Technical Field

[0001] The present invention relates to the technical field of track prediction, and in particular to a track prediction intelligent optimization system and method. Background Art

[0002] In the aviation field, accurately predicting the trajectory of an aircraft is very important for ensuring flight safety and improving traffic efficiency. However, due to the complexity and uncertainty of air traffic, traditional prediction methods often have certain limitations. For trajectory prediction, existing technologies mainly include prediction based on physical models, prediction based on data-driven, and hybrid methods. These methods have their own advantages, but there are also some problems.

[0003] The prediction method based on physical models mainly uses the dynamic and kinematic equations of spacecraft to estimate the trajectory. The advantages of this method are high accuracy and strong interpretability, but it requires accurate model parameters and initial conditions. However, in practical applications, model parameters are often difficult to obtain accurately, and the movement of spacecraft is affected by many factors, such as atmospheric drag, solar radiation pressure, etc., which are difficult to fully consider in physical models.

[0004] Data-driven prediction methods rely on a large amount of historical data to train the model so that it can learn the laws of spacecraft motion. The advantage of this method is that it does not require an explicit physical model and can adapt to complex nonlinear relationships. However, it also has some problems, such as strong data dependence, sensitivity to outliers, and limited generalization ability. In addition, for new spacecraft or new mission scenarios, there may be a lack of sufficient historical data to train the model.

[0005] The hybrid method attempts to combine the advantages of physical models and data-driven methods to improve the accuracy and robustness of trajectory prediction. However, this method also faces some challenges, such as how to effectively integrate physical models and data-driven information, and how to deal with the uncertainty between models and data.

[0006] In general, existing trajectory prediction technologies have certain limitations and problems. Existing prediction methods give little consideration to the selection of features related to the time dimension and spatial latitude, resulting in inaccurate prediction accuracy.

[0007] Therefore, a track prediction intelligent optimization system and method are proposed. Summary of the invention

[0008] The purpose of the present invention is to provide an intelligent optimization system and method for track prediction, so as to solve the problem that the existing prediction methods proposed in the above background technology give little consideration to the selection of time dimension and space latitude related features, resulting in inaccurate prediction accuracy.

[0009] To achieve the above object, the present invention provides the following technical solutions: a track prediction intelligent optimization system and method, comprising the following steps;

[0010] Step 1: Preprocess the TLE data and obtain the data set;

[0011] Step 2: Train the dataset from step 1 using a machine learning model to find possible behavior patterns of satellites at different time points.

[0012] Step 3: Propagate the current state through the aircraft dynamics using an estimation algorithm, and then correct the propagated state using pseudo-measurement methods to take into account the expected future behavior and the current motion of the aircraft;

[0013] Step 4: Model evaluation and optimization;

[0014] Step 5: Prediction and classification.

[0015] Preferably, the TLE data in step one is obtained from content published by a professional aerospace agency, and the obtained TLE data includes the position parameters and speed parameters of the satellite. The TLE data is parsed into a usable numerical format through HMMS, and the satellite's position data information and speed data information are extracted from the parsed data as an observation state sequence.

[0016] Preferably, the machine learning model training in step 2 specifically includes the following three steps:

[0017] S1. Select and determine the hidden state sequence probability and the observed state sequence probability;

[0018] S2, initialize the parameters of HMMS, and use the Baum-Welch algorithm to estimate the parameters of HMMS;

[0019] S3. Use the Viterbi algorithm to find the most accurate hidden state sequence of the parameters of HMMS after Baum-Welch estimation, which helps to find the possible behavior patterns of satellites at different time points.

[0020] Preferably, the hidden state sequence probability is represented by a matrix A, where A[i,j] represents the probability of transitioning from state i to state j;

[0021] The observed state sequence probability is represented by a matrix B, where B[j,k] represents the probability of generating an observation value k in state j.

[0022] Preferably, the Baum-Welch estimation process is as follows:

[0023] HMMS parameters include the state transition probability matrix A[i,j], the observation probability matrix B[j,k], and the initial state probability distribution π = {π i}, through the given observation sequence ο = {ο1,ο2,···,ο T Estimate the position data and velocity data extracted by HMMS to maximize the likelihood of the observation sequence; and optimize the parameters step by step through iteration using the EM algorithm. Each iteration is divided into E steps and M steps, where:

[0024] Step E: Calculate the forward variable α t (i), α t (i) = P(ο1,ο2,···,ο t ,q t =iλ), where time t is in state i, and λ = (π, A, B) is the position data information and speed data information extracted by HMMS;

[0025] Calculate the backward variable β t (i), β t (i) = P(ο t+1 ,ο t+2 ,···,ο t |q t =i,λ);

[0026] Calculate the probability of being in state i at time t and in state j at time t+1, the formula is as follows;

[0027]

[0028] Where N is the total number of states;

[0029] Calculate the probability of being in state i at time t, the formula is as follows;

[0030]

[0031] Step M: Update the state transition probability matrix A:

[0032] Update the observation probability matrix B:

[0033] Update the initial state probability distribution π: π i =ε1(i);

[0034] Repeat the above steps E and M until the likelihood converges or the preset number of iterations is reached to estimate the parameters of HMMS, and obtain the parameter state set of HMMS as S = {s1, s2, ···, s N}.

[0035] Preferably, the specific formula of the Viterbi algorithm is as follows:

[0036] The state set of HMMS is S = {s1,s2,···,s N}, the observation sequence is ο={ο1,ο2,···,ο T}, the state transition probability matrix is ​​A[i,j], and the observation probability matrix is ​​B[j,k];

[0037] Define δ t (i) is the maximum probability of reaching state i at time t, and ψ is recorded at the same time t (i) is the state before reaching state i at time t;

[0038] Initialization: δ1(i)=π(i)b[i,o1], where π is the initial state probability, i=1,2,...,N;

[0039] Recursion: For t=2,3...,T and i=1,2,...,N;

[0040]

[0041] termination:

[0042]

[0043] Backtrack:

[0044] For t = T-1, T-2, ..., 1:

[0045]

[0046] Finally, the relatively accurate hidden state sequence is obtained:

[0047] Preferably, the specific method of step three is that the trained model generates a pseudo measurement by taking the previous measurement to the time step t, the estimation algorithm UKF propagates the current state through the aircraft dynamics, and then uses the pseudo measurement method to correct the propagation state, and uses the HMMS to predict the data-driven pseudo measurement value, that is, N((t+1), (t+1)), as the input of the posterior prediction, and the specific formula is as follows;

[0048]

[0049] Among them, S K+1 represents the measurement covariance matrix at K+1, R represents the statistical characteristics of the time dimension and space dimension during the measurement process, and H K+1 The Jacobian matrix representing the measurement function is used to describe the linear relationship between the measurement value and the state, P K+1|Krepresents the covariance matrix of the predicted state at time K+1 based on time K, Indicates H K+1 The transposed matrix of , δ represents the weight coefficient, the weight coefficient value is 1, N((t+1), (t+1)) represents the pseudo measurement value driven by HMMS prediction data,

[0050] N((t+1),(t+1)) T Represents the transposed matrix of N((t+1),(t+1)).

[0051] Compared with the prior art, the present invention has the following beneficial effects:

[0052] 1. In the present invention, a flight prediction framework based on HMMS and UKF is proposed. The framework trains a machine learning-based trajectory prediction model using historical monitoring data to represent the collective behavior of a set of flight trajectories, and integrates the machine learning predictions as pseudo-measurements through a physical estimation algorithm to take into account the current motion of the aircraft. The framework is tested using real air traffic monitoring data in real-time air traffic control applications to predict the future state information of the flight.

[0053] 2. In the present invention, a trajectory prediction method based on HMMS is used for trajectory prediction. The trajectory prediction method based on HMMS can capture the complex changes of flight trajectories in time and space dimensions. In air traffic surveillance data, flight trajectories are expressed as time series data. By using HMMS, we can regard these trajectories as observation sequences generated by a series of hidden states. Each hidden state corresponds to a specific flight mode. The transition probabilities between these hidden states and the probabilities of them generating observation sequences can be modeled and learned through HMMS. This method not only takes into account the similarity of trajectories in the spatial dimension, but also effectively captures the dynamic changes in the time dimension, thereby more accurately predicting flight trajectories.

[0054] 3. In the present invention, machine learning prediction is used as pseudo-measurement prediction based on UKF. The current state is propagated through the aircraft dynamics model, and then the propagated state is corrected using pseudo-measurement data. The pseudo-measurement value driven by HMMS prediction data, i.e., N((t+1), (t+1)), is used as the input of the posterior prediction, which can effectively combine the prediction ability of machine learning and realize accurate estimation of the future state of the aircraft. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 It is a schematic diagram of the process of the present invention;

[0056] Figure 2 Schematic diagram of the process of the machine learning model of the present invention. DETAILED DESCRIPTION

[0057] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0058] See also Figure 1 and Figure 2 The present invention provides a technical solution of a track prediction intelligent optimization system and method:

[0059] A track prediction intelligent optimization system and method, comprising the following steps:

[0060] Step 1: Preprocess the TLE data and obtain a data set. The TLE data in step 1 is obtained from the content released by a professional aerospace agency. The obtained TLE data includes the position parameters and speed parameters of the satellite. The TLE data is parsed into a usable numerical format through HMMS, and the satellite position data information and speed data information are extracted from the parsed data as an observation state sequence;

[0061] Step 2: Train the data set in step 1 by using a machine learning model to find possible behavior patterns of satellites at different time points, specifically including the following steps;

[0062] S1. Select and determine the hidden state sequence probability and the observed state sequence probability;

[0063] Among them, the hidden state sequence probability represents the probability of selecting a stable orbit state, which means that the satellite is operating stably on its predetermined orbit without major orbit adjustments or maneuvers, and the orbital parameters of the satellite will remain relatively stable under this state probability. The hidden state sequence probability is represented by the matrix A, where A[i,j] represents the probability of transitioning from state i to state j;

[0064] Among them, the observed state sequence probability represents the position, velocity and acceleration of the selected satellite, and the observed state sequence probability is represented by a matrix B, where B[j,k] represents the probability of generating an observation value k in state j.

[0065] S2, initialize the parameters of HMMS, and use the Baum-Welch algorithm to estimate the parameters of HMMS;

[0066] The Baum-Welch algorithm is a commonly used algorithm for training HMMS. It is a special case of the expectation-maximization EM algorithm. The estimation process of Baum-Welch is as follows:

[0067] HMMS parameters include the state transition probability matrix A[i,j], the observation probability matrix B[j,k], and the initial state probability distribution π = {π i}, through the given observation sequence ο = {ο1,ο2,···,ο T Estimate the position data and velocity data information extracted by HMMS to maximize the likelihood probability of the observation sequence; and optimize the parameters step by step in an iterative manner through the EM algorithm. Each iteration is divided into E steps and M steps, where:

[0068] Step E: Calculate the forward variable α t (i), α t (i) = P(ο1,ο2,···,ο t ,q t =i|λ), where time t is in state i, and λ = (π, A, B) is the position data information and speed data information extracted by HMMS;

[0069] Calculate the backward variable β t (i), β t (i) = P(ο t+1 ,ο t+2 ,···,ο t |q t =i,λ);

[0070] Calculate the probability of being in state i at time t and in state j at time t+1, the formula is as follows;

[0071]

[0072] Where N is the total number of states;

[0073] Calculate the probability of being in state i at time t, the formula is as follows;

[0074]

[0075] Step M: Update the state transition probability matrix A:

[0076] Update the observation probability matrix B:

[0077] Update the initial state probability distribution π: π i =ε1(i);

[0078] Repeat the above steps E and M until the likelihood converges or the preset number of iterations is reached to estimate the parameters of HMMS, and obtain the parameter state set of HMMS as S = {s1, s2, ···, s N}.

[0079] S3, using the Viterbi algorithm to find the most accurate hidden state sequence of the parameters of the HMMS after Baum-Welch estimation, to help find the possible behavior patterns of the satellite at different time points;

[0080] The specific formula of the Viterbi algorithm is as follows:

[0081] The state set of HMMS is S = {s1,s2,···,s N}, the observation sequence is ο={ο1,ο2,···,ο T}, the state transition probability matrix is ​​A[i,j], and the observation probability matrix is ​​B[j,k];

[0082] Define δ t (i) is the maximum probability of reaching state i at time t, and ψ is recorded at the same time t (i) is the state before reaching state i at time t;

[0083] Initialization: δ1(i)=π(i)b[i,o1], where π is the initial state probability, i=1,2,...,N;

[0084] Recursion: For t=2,3...,T and i=1,2,...,N;

[0085]

[0086] termination:

[0087]

[0088] Backtrack:

[0089] For t = T-1, T-2, ..., 1:

[0090]

[0091] Finally, the relatively accurate hidden state sequence is obtained:

[0092] Step 3: Use the estimation algorithm to propagate the current state through aircraft dynamics, and then use the pseudo-measurement method to correct the propagated state. The specific method of step 3 is that the trained model generates a pseudo-measurement by taking the previous measurement to the time step t. The estimation algorithm UKF propagates the current state through aircraft dynamics, and then uses the pseudo-measurement method to correct the propagated state. The HMMS predicts the data-driven pseudo-measurement value, that is, N((t+1), (t+1)), as the input of the posterior prediction. The specific formula is as follows;

[0093]

[0094] Among them, S K+1 represents the measurement covariance matrix at K+1, R represents the statistical characteristics of the time dimension and space dimension during the measurement process, and H K+1 The Jacobian matrix representing the measurement function is used to describe the linear relationship between the measurement value and the state, P K+1K represents the covariance matrix of the predicted state at time K+1 based on time K, Indicates H K+1 The transposed matrix of , δ represents the weight coefficient, the weight coefficient value is 1, N((t+1), (t+1)) represents the pseudo measurement value driven by HMMS prediction data,

[0095] N((t+1),(t+1)) T represents the transposed matrix of N((t+1),(t+1));

[0096] Step 4: Model evaluation and optimization;

[0097] Step 5: Prediction and classification.

[0098] At work, a flight prediction framework based on HMMS and UKF is proposed. The framework trains a machine learning-based trajectory prediction model using historical monitoring data to represent the collective behavior of a set of flight trajectories, and integrates the machine learning predictions as pseudo-measurements through a physical estimation algorithm to take into account the current motion of the aircraft. The framework is tested using real air traffic monitoring data in real-time air traffic control applications to predict the future state information of the flight. The trajectory prediction method based on HMMS is used for trajectory prediction. The trajectory prediction method based on HMMS is able to capture the complex changes of flight trajectories in time and space dimensions. In air traffic monitoring data, flight trajectories are presented as time series data. By utilizing HMMS, we can view these trajectories as observation sequences generated by a series of hidden states, each of which corresponds to a specific flight mode. The transition probabilities between these hidden states and the probabilities of them generating observation sequences can be modeled and learned through HMMS. This approach not only takes into account the similarity of trajectories in the spatial dimension, but also effectively captures the dynamic changes in the time dimension, thereby more accurately predicting flight trajectories, using machine learning predictions as pseudo-measurement predictions based on UKF (Unscented Kalman Filter) to comprehensively consider the expected future behavior and the current motion of the aircraft. The trained model generates a pseudo-measurement by processing the previous measurement data to time step t, propagates the current state through the aircraft dynamics model, and then uses the pseudo-measurement data to correct the propagated state. Specifically, this method uses the pseudo-measurement driven by the HMMS prediction data, i.e., N((t+1),(t+1)), as the input of the (posterior) prediction. In this way, UKF can effectively combine the predictive capabilities of machine learning to achieve an accurate estimation of the future state of the aircraft.

[0099] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A track prediction intelligent optimization system and method, characterized by: The steps include: Step 1: Preprocess the TLE data and obtain the data set; Step 2: Train the dataset from step 1 using a machine learning model to find possible behavior patterns of satellites at different time points. Step 3: Propagate the current state through the aircraft dynamics using an estimation algorithm, and then correct the propagated state using pseudo-measurement methods to take into account the expected future behavior and the current motion of the aircraft; Step 4: Model evaluation and optimization; Step 5: Prediction and classification.

2. A track prediction intelligent optimization system and method according to claim 1, characterized in that: The TLE data includes the position parameters and speed parameters of the satellite. The TLE data is parsed into a usable numerical format through HMMS, and the satellite's position data information and speed data information are extracted from the parsed data as an observation state sequence.

3. The track prediction intelligent optimization system and method according to claim 1, characterized in that: The machine learning model training in step 2 specifically includes the following steps: S1. Select and determine the hidden state sequence probability and the observed state sequence probability; S2, initialize the parameters of HMMS, and use the Baum-Welch algorithm to estimate the parameters of HMMS; S3. Use the Viterbi algorithm to find the most accurate hidden state sequence of the parameters of HMMS after Baum-Welch estimation, which helps to find the possible behavior patterns of satellites at different time points.

4. A track prediction intelligent optimization system and method according to claim 3, characterized in that: The hidden state sequence probability is represented by a matrix A, where A[i,j] represents the probability of transitioning from state i to state j; The observed state sequence probability is represented by a matrix B, where B[j,k] represents the probability of generating an observation value k in state j.

5. The track prediction intelligent optimization system and method according to claim 3, characterized in that: The Baum-Welch estimation process is as follows: HMMS parameters include the state transition probability matrix A[i,j], the observation probability matrix B[j,k], and the initial state probability distribution π = {π i }, through the given observation sequence ο = {ο1,ο2,···,ο T Estimate the position data and velocity data information extracted by HMMS to maximize the likelihood probability of the observation sequence; and optimize the parameters step by step in an iterative manner through the EM algorithm. Each iteration is divided into E steps and M steps, where: Step E: Calculate the forward variable α t (i), α t (i) = P(ο1,ο2,···,ο t ,q t =iλ), where time t is in state i, and λ = (π, A, B) is the position data information and speed data information extracted by HMMS; Calculation backward change amount β t (i), β t (i)=P(ο t+1 ,ο t+2 , ,ο t |q t = i, λ); Calculate the probability of being in state i at time t and in state j at time t+1, the formula is as follows; Where N is the total number of states; Calculate the probability of being in state i at time t, the formula is as follows; Step M: Update the state transition probability matrix A: Update the observation probability matrix B: Update the initial state probability distribution π: π i =ε1(i); Repeat the above steps E and M until the likelihood converges or the preset number of iterations is reached to estimate the parameters of HMMS, and obtain the parameter state set of HMMS as S = {s1, s2, ···, s N }.

6. A track prediction intelligent optimization system and method according to claim 5, characterized in that: The specific formula of the Viterbi algorithm is as follows: The state set of HMMS is S = {s1,s2,···,s N }, the observation sequence is ο={ο1,ο2,···,ο T }, the state transition probability matrix is ​​A[i,j], and the observation probability matrix is ​​B[j,k]; Definition of δ t (i) is the maximum probability of reaching state i at time t, and ψ is recorded at the same time t (i) is the state before reaching state i at time t; Initialization: δ1(i)=π(i)b[i,o1], where π is the initial state probability, i=1,2,...,N; Recursion: For t=2,3...,T and i=1,2,...,N; termination: Backtrack: For t = T-1, T-2, ..., 1: Finally, the relatively accurate hidden state sequence is obtained:

7. The track prediction intelligent optimization system and method according to claim 1, characterized in that: The specific method of step three is that the trained model generates a pseudo measurement by taking the previous measurement to the time step t, the estimation algorithm UKF propagates the current state through the aircraft dynamics, and then uses the pseudo measurement method to correct the propagated state, and uses HMMS to predict the data-driven pseudo measurement value, that is, N((t+1), (t+1)), as the input of the posterior prediction. The specific formula is as follows; Among them, S K+1 represents the measurement covariance matrix at K+1, R represents the statistical characteristics of the time dimension and space dimension during the measurement process, and H K+1 The Jacobian matrix representing the measurement function is used to describe the linear relationship between the measurement value and the state, P K+1K represents the covariance matrix of the predicted state at time K+1 based on time K, Indicates H K+1 The transposed matrix of , δ represents the weight coefficient, the weight coefficient value is 1, N((t+1), (t+1)) represents the pseudo measurement value driven by HMMS prediction data, N((t+1),(t+1)) T Represents the transposed matrix of N((t+1),(t+1)).

Images