A method for aircraft state estimation combining multi-model interactive filtering and DNN

Abstract

The present application belongs to the technical field of aircraft state estimation. The present application provides an aircraft state estimation method combining multi-model interactive filtering and DNN. The disclosed embodiment trains the LSTM network model according to the collected aircraft trajectory set, and the LSTM-KF filter uses the nonlinear function formed by the LSTM network model to replace the original state transfer matrix for state prediction update, and connects the LSTM-KF filter as a filter module to the interactive multi-model filtering framework. This method has interactions between multiple different models, and can adapt well to the complex motion mode of the aircraft, even when the aircraft motion mode jumps, it can still adapt well, and at the same time integrates the results of the deep neural network filter to improve the upper limit of the aircraft state estimation accuracy.

Description

Technical Field

[0001] The disclosed embodiments relate to the technical field of aircraft state estimation, and in particular to an aircraft state estimation method combining multi-model interactive filtering and DNN. Background Art

[0002] Aircraft trajectory estimation is of great significance in improving flight safety, flight efficiency, and supporting the execution of complex tasks. It is one of the basic tasks in the field of aerospace. For aircraft with complex maneuvers, their trajectories are more complex than those of general aircraft. Most of the current aircraft state estimation methods use models such as constant velocity (CV), constant acceleration (CA), constant turn (CT), Singer, etc. to describe target maneuvers, and implement state estimation based on methods such as extended Kalman filter (EKF) and unscented Kalman filter (UKF). When faced with such aircraft trajectory estimation problems with complex maneuvers, limited by the accuracy of the model, traditional state estimation methods cannot fully adapt to the complex motion modes of the aircraft, resulting in low accuracy of state estimation. Compared with the traditional Kalman filter method, the state estimation accuracy of the deep learning method is higher and is not limited by the accuracy of the model, but the deep learning method has certain instability, including overfitting problems.

[0003] Therefore, it is necessary to improve one or more problems existing in the above-mentioned related technical solutions.

[0004] It should be noted that this section is intended to provide background or context for the technical solutions of the present disclosure stated in the claims. The description herein is not admitted to be prior art by virtue of being included in this section. Summary of the invention

[0005] In order to avoid the shortcomings of the prior art, the present application provides an aircraft state estimation method that combines multi-model interactive filtering and DNN to solve the instability problems such as overfitting in the prior art.

[0006] According to an embodiment of the present disclosure, a method for aircraft state estimation combining multi-model interactive filtering and DNN is provided, the method comprising:

[0007] Determine a number of kinematic models according to the state variables of the aircraft, and establish a number of corresponding traditional kinematic Kalman filters according to the number of kinematic models; wherein the kinematic models include a uniform speed model, a uniform acceleration model and a uniform speed turning model;

[0008] Build and train the LSTM network model, and build the LSTM-KF filter based on the trained LSTM network model;

[0009] Build an interactive multi-model filtering framework and add all traditional kinematic Kalman filters and LSTM-KF filters to the interactive multi-model filtering framework;

[0010] The historical position information is input into the interactive multi-model filtering framework, all traditional kinematic Kalman filters and LSTM-KF filters are traversed for prediction updates, the model probability is calculated and updated, and the final state estimation result is obtained through weighted calculation.

[0011] Furthermore, the state variables include:

[0012] The three-axis position, three-axis velocity, three-axis acceleration and turning rate of the aircraft; among which,

[0013] The state variables of the uniform speed model are the three-axis position and three-axis velocity of the aircraft. The state variables of the uniform acceleration model are the three-axis position, three-axis velocity and three-axis acceleration of the aircraft. The state variables of the uniform turning model include the three-axis position, three-axis velocity and turning rate of the aircraft. The state variables of the LSTM network model are the three-axis position and three-axis velocity of the aircraft.

[0014] Furthermore, the steps of constructing and training the LSTM network model and constructing the LSTM-KF filter according to the trained LSTM network model include:

[0015] Obtain the position information sequence of the aircraft, take the first M frames of the position information sequence as input and the M+1th frame as the true value to form a data set;

[0016] Divide the data set into training set and test set according to the preset ratio;

[0017] Build LSTM network model;

[0018] Use the training set to train the LSTM network model, and use the test set to test the trained LSTM network model to obtain a trained LSTM network model;

[0019] The nonlinear function of the trained LSTM network model is used instead of the state transfer matrix to update the state and form an LSTM-KF filter.

[0020] Furthermore, the input of the LSTM network model is the historical trajectory time series information, and the output is the position prediction value at the current moment, that is:

[0021]

[0022] in, is the time step The predicted value of the LSTM network model, is the inference function of the LSTM network model, is the time step The estimated value of the position, is the time step The estimated value of the position, is the time step The estimated value of the position of

[0023] The specific structure of the LSTM network model is:

[0024]

[0025]

[0026]

[0027]

[0028]

[0029]

[0030] in, is the output of the forget gate at time t, is the sigmoid activation function, To act on the hidden state The weight matrix of To act on the current input The weight matrix of is the hidden state at time t-1, is the hidden state at time t, is the input of the network at each moment, is the state of the network cell at time t, is the bias vector of the forget gate, is the output of the input gate at time t, To act on the hidden state The weight matrix of To act on input The weight matrix of is the bias vector of the input gate, is the output of the output gate at time t, To act on the hidden state The weight matrix of To act on input The weight matrix of is the bias vector of the output gate, is the candidate cell state, is the tanh activation function, To act on the current input The weight matrix of To act on input The weight matrix of is the bias vector of the candidate cell state, is the network cell state at time t-1, Represents element-wise multiplication.

[0031] Furthermore, the historical position information is input into the interactive multi-model filtering framework, all traditional kinematic Kalman filters and LSTM-KF filters are traversed for prediction update, the model probability is calculated and updated, and the final state estimation result is obtained through weighted calculation, including:

[0032] The Markov transition probability matrix is ​​used to realize the model interaction between the traditional kinematic Kalman filter and the LSTM-KF filter. The mixed estimate and covariance are calculated according to the state estimate and model probability of each traditional kinematic Kalman filter at the previous moment, and the mixed estimate is used as the initial state of the current cycle.

[0033] The results of the hybrid estimation are input into all traditional kinematic Kalman filters and LSTM-KF filters for prediction and update steps;

[0034] The predicted state obtained by each traditional kinematic Kalman filter is combined with the observed value to calculate the observation likelihood, the weighted likelihood of each traditional kinematic Kalman filter is calculated in combination with the model weight, and the sum of all weighted likelihoods is calculated as a normalization factor and the model probability is updated;

[0035] Based on the model probability, the estimation results of each traditional kinematic Kalman filter are weighted and combined to obtain the final state estimation and covariance estimation.

[0036] Furthermore, the prediction steps of the LSTM-KF filter are:

[0037]

[0038]

[0039] in, is for The Jacobian matrix of is the system noise matrix, is the time step The predicted covariance matrix;

[0040] The update steps of LSTM-KF filter are:

[0041]

[0042]

[0043]

[0044] in, is the Kalman gain, is the observation matrix, is the observation noise matrix, is at the time step The observed value of is the time step The final estimate of is the time step The covariance matrix of It is a unit array.

[0045] Furthermore, the Markov transition probability matrix is:

[0046]

[0047] in, is the Markov transition probability matrix.

[0048] Further, the steps of calculating the observation likelihood according to the predicted state obtained by each traditional kinematic Kalman filter in combination with the observed value, calculating the weighted likelihood of each traditional kinematic Kalman filter in combination with the model weight, and calculating the sum of all weighted likelihoods as a normalization factor and updating the model probability include:

[0049] Calculate the j The likelihood function of the model is:

[0050]

[0051] in, is the measurement error covariance matrix, is the measurement error;

[0052] Calculate the normalization constant:

[0053]

[0054] in, is the credibility of the fused model;

[0055] Calculate the j Update model probability of models:

[0056]

[0057] in, is the model probability at the k-1th moment.

[0058] Furthermore, based on the model probability, the estimation results of each traditional kinematic Kalman filter are weighted and combined to obtain the final state estimation and covariance estimation steps, including:

[0059]

[0060]

[0061] in, is the final state estimation result, For the j Model The estimated position at time, is the updated Markov transition probability matrix, is the covariance matrix of the j-th model.

[0062] The technical solution provided by the embodiments of the present disclosure may have the following beneficial effects:

[0063] In the embodiment of the present disclosure, through the above-mentioned joint multi-model interactive filtering and DNN aircraft state estimation method, on the one hand, the LSTM network model is trained according to the collected aircraft trajectory set, and the LSTM-KF filter uses the nonlinear function composed of the LSTM network model to replace the original state transfer matrix for state prediction update, and the LSTM-KF filter is connected to the interactive multi-model filtering framework as a filter module. This method has interactions between multiple different models, and can adapt well to the complex motion mode of the aircraft, even when the aircraft motion mode jumps, it can still adapt well, and at the same time integrates the results of the deep neural network filter, which improves the upper limit of the accuracy of the aircraft state estimation. On the other hand, the LSTM-KF filter is used to improve the upper limit of the accuracy of the interactive multi-model filter, and when the deep neural network prediction results are biased at certain moments, it can be switched to the traditional kinematics-based filter in time, ensuring the system performance of the interactive multi-model filter. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] The accompanying drawings herein are incorporated into the specification and constitute a part of the specification, illustrate embodiments consistent with the present disclosure, and together with the specification are used to explain the principles of the present disclosure. Obviously, the accompanying drawings described below are only some embodiments of the present disclosure, and for ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without creative work.

[0065] Figure 1 A step diagram showing a method for estimating an aircraft state by combining multi-model interactive filtering and DNN in an exemplary embodiment of the present disclosure;

[0066] Figure 2 A schematic diagram showing a method for estimating an aircraft state by combining multi-model interactive filtering and DNN in an exemplary embodiment of the present disclosure is shown;

[0067] Figure 3 A schematic diagram of an LSTM-KF filter in an exemplary embodiment of the present disclosure is shown;

[0068] Figure 4 The weight proportion of each model when estimating the actual aircraft state in the exemplary embodiment of the present disclosure is shown;

[0069] Figure 5 The filtering result of the actual aircraft state estimation in an exemplary embodiment of the present disclosure is shown. DETAILED DESCRIPTION

[0070] Example embodiments will now be described more fully with reference to the accompanying drawings. However, example embodiments can be implemented in a variety of forms and should not be construed as limited to the examples set forth herein; rather, these embodiments are provided so that the disclosure will be more comprehensive and complete and to fully convey the concepts of the example embodiments to those skilled in the art. The described features, structures, or characteristics may be combined in any suitable manner in one or more embodiments.

[0071] In addition, the accompanying drawings are only schematic illustrations of the embodiments of the present disclosure and are not necessarily drawn to scale. The same reference numerals in the figures represent the same or similar parts, and their repeated descriptions will be omitted. Some of the block diagrams shown in the accompanying drawings are functional entities and do not necessarily correspond to physically or logically independent entities.

[0072] This example embodiment provides an aircraft state estimation method that combines multi-model interactive filtering and DNN. Figure 1 As shown in , the aircraft state estimation method of the joint multi-model interactive filtering and DNN may include: step S101 to step S104.

[0073] Step S101: determining a plurality of kinematic models according to the state variables of the aircraft, and establishing a plurality of corresponding traditional kinematic Kalman filters according to the plurality of kinematic models; wherein the kinematic models include a uniform speed model, a uniform acceleration model and a uniform speed turning model;

[0074] Step S102: construct and train an LSTM network model, and construct an LSTM-KF filter based on the trained LSTM network model;

[0075] Step S103: construct an interactive multi-model filtering framework, and add all traditional kinematic Kalman filters and LSTM-KF filters to the interactive multi-model filtering framework;

[0076] Step S104: Input the historical position information into the interactive multi-model filtering framework, traverse all traditional kinematic Kalman filters and LSTM-KF filters for prediction update, calculate and update the model probability, and obtain the final state estimation result through weighted calculation.

[0077] Through the above-mentioned aircraft state estimation method of joint multi-model interactive filtering and DNN, on the one hand, the LSTM network model is trained according to the collected aircraft trajectory set, and the LSTM-KF filter uses the nonlinear function composed of the LSTM network model to replace the original state transfer matrix for state prediction update, and the LSTM-KF filter is connected to the interactive multi-model filtering framework as a filter module. This method has interactions between multiple different models and can adapt well to the complex motion mode of the aircraft, even when the aircraft motion mode jumps. It can still adapt well, and at the same time integrates the results of the deep neural network filter to improve the upper limit of the aircraft state estimation accuracy. On the other hand, the use of LSTM-KF filter improves the upper limit of the accuracy of the interactive multi-model filter. When the deep neural network prediction results are biased at certain moments, it can switch to the traditional kinematics-based filter in time, ensuring the system performance of the interactive multi-model filter.

[0078] Next, we will refer to Figures 1 to 5 Each step of the above-mentioned joint multi-model interactive filtering and DNN aircraft state estimation method in this example implementation is described in more detail.

[0079] In step S101, Figure 2 As shown, it is a schematic diagram of the aircraft state estimation method combining multi-model interactive filtering and DNN.

[0080] The traditional kinematic Kalman filter includes Kalman filters of the uniform velocity (CV) model, uniform acceleration (CA) model, and uniform turning (CT) model. Among them, the state variables of the uniform velocity model are the three-axis position and three-axis velocity of the aircraft, the state variables of the uniform acceleration model are the three-axis position, three-axis velocity, and three-axis acceleration of the aircraft, the state variables of the uniform turning model include the three-axis position, three-axis velocity, and turning rate of the aircraft, and the state variables of the LSTM network model are the three-axis position and three-axis velocity of the aircraft. The state variables of the entire interactive multi-model filter are selected as the three-axis position, three-axis velocity, three-axis acceleration, and turning rate of the aircraft.

[0081] In step S102, an aircraft simulation system is constructed, and 5,000 aircraft trajectories are simulated to form a trajectory set. Sequence information of the aircraft position is intercepted on each trajectory, and the previous frame is used as input and the first frame is used as the true value to form a data set. 80% of the data set is used as a training set and 20% as a test set. The network will be trained using the training set using the Adam optimizer in an end-to-end manner. The nonlinear function of the trained LSTM model is used instead of the state transfer matrix to update the state to form an LSTM-KF filter. Figure 3 As shown, this is the schematic diagram of the LSTM-KF filter.

[0082] The steps of constructing and training the LSTM network model and constructing the LSTM-KF filter according to the trained LSTM network model include:

[0083] Obtain the position information sequence of the aircraft, take the first M frames of the position information sequence as input and the M+1th frame as the true value to form a data set;

[0084] Divide the data set into training set and test set according to the preset ratio;

[0085] Build LSTM network model;

[0086] Use the training set to train the LSTM network model, and use the test set to test the trained LSTM network model to obtain a trained LSTM network model;

[0087] The nonlinear function of the trained LSTM network model is used instead of the state transfer matrix to update the state and form an LSTM-KF filter.

[0088] In addition, the input of the LSTM network model is the historical trajectory time series information, and the output is the position prediction value at the current moment, that is:

[0089]

[0090] in, is the time step The predicted value of the LSTM network model, is the inference function of the LSTM network model, is the time step The estimated value of the position, is the time step The estimated value of the position, is the time step The estimated value of the position of

[0091] The specific structure of the LSTM network model is:

[0092]

[0093] in, is the output of the forget gate at time t, is the sigmoid activation function, To act on the hidden state The weight matrix of To act on the current input The weight matrix of is the hidden state at time t-1, is the hidden state at time t, is the input of the network at each moment, is the state of the network cell at time t, is the bias vector of the forget gate, is the output of the input gate at time t, To act on the hidden state The weight matrix of To act on input The weight matrix of is the bias vector of the input gate, is the output of the output gate at time t, To act on the hidden state The weight matrix of To act on input The weight matrix of is the bias vector of the output gate, is the candidate cell state, is the tanh activation function, To act on the current input The weight matrix of To act on input The weight matrix of is the bias vector of the candidate cell state, is the network cell state at time t-1, Represents element-wise multiplication.

[0094] In step S103, the model interaction between the traditional kinematic Kalman filter and the LSTM-KF filter is realized by using the Markov transition probability matrix, and the mixed estimate and covariance are calculated according to the state estimate and model probability of each traditional kinematic Kalman filter at the previous moment, and the mixed estimate is used as the initial state of the current cycle; the result of the mixed estimate is input into all traditional kinematic Kalman filters and LSTM-KF filters for prediction and update steps; the observation likelihood is calculated according to the predicted state obtained by each traditional kinematic Kalman filter combined with the observed value, and the weighted likelihood of each traditional kinematic Kalman filter is calculated in combination with the model weight, and the sum of all weighted likelihoods is calculated as a normalization factor and the model probability is updated; based on the model probability, the estimation results of each traditional kinematic Kalman filter are weighted and merged to obtain the final state estimate and covariance estimate. Among them, the traditional kinematic model Kalman filter predicts according to the estimated value at the previous moment, and the LSTM-KF filter predicts according to the estimated value at the previous M moments.

[0095] Specifically, we use Windows system, Visual Studio Code platform, and Python 3.8 as programming language to build an interactive multi-model filtering framework, and connect the Kalman filter and LSTM-KF filter of the uniform speed model, uniform acceleration model, and uniform turning model.

[0096] The global state variables include the state variables shared by each model and the state variables unique to each model, but in a single filter, only the local state variables are updated. The CV model only updates the position and velocity state variables, the CA model updates the position, velocity, and acceleration state variables, the CT model updates the position, velocity, and turning rate state variables, and the LSTM-KF model updates the position and velocity state variables, which reduces the computational complexity and improves the efficiency of the algorithm.

[0097] The main idea of ​​the LSTM-KF filter is to use the Kalman filter without specifying the linear state transfer matrix, but to predict it using the nonlinear function formed by the LSTM network, so that the model can learn rich and dynamic Kalman filter components from the data, such as Figure 3 shown.

[0098] The prediction steps of the LSTM-KF filter are:

[0099]

[0100]

[0101] in, is for The Jacobian matrix of is the system noise matrix, is the time step The predicted covariance matrix;

[0102] The update steps of LSTM-KF filter are:

[0103]

[0104]

[0105]

[0106] in, is the Kalman gain, is the observation matrix, is the observation noise matrix, is at the time step The observed value of is the time step The final estimate of is the time step The covariance matrix of It is a unit array.

[0107] The Markov transition probability matrix is ​​used to realize data input interaction. The transition probability from model i to model j is, then the Markov transition probability matrix is:

[0108]

[0109] in, is the Markov transition probability matrix.

[0110] The model probability update process is as follows:

[0111] Calculate the j The likelihood function of the model is:

[0112]

[0113] in, is the measurement error covariance matrix, is the measurement error;

[0114] Calculate the normalization constant:

[0115]

[0116] in, is the credibility of the fused model;

[0117] Calculate the j Update model probability of models:

[0118]

[0119] in, is the model probability at the k-1th moment.

[0120] In step S104, historical position information is input, each filter is traversed to perform prediction update, model probability is calculated and updated, and the final state estimation result is obtained through weighted calculation, and the above process is repeated continuously.

[0121] The preprocessing of time series information includes two parts. For the traditional kinematic Kalman filter, the update of the current state only requires the state of the previous state to be multiplied by the state transfer matrix. For the LSTM-KF filter, the input is a time series of length N for prediction to ensure the accuracy of the LSTM-KF filter prediction.

[0122] The mixing probability is calculated based on the state probability distribution and the model weight, and then the mixed state is calculated based on the input state and the mixing probability. The mixed state is input to each filter module for the Kalman filter prediction and update steps.

[0123] The global state variables include the state variables shared by each model and the state variables unique to each model, but in a single filter, only the local state variables are updated. The CV model only updates the position and velocity state variables, the CA model updates the position, velocity, and acceleration state variables, the CT model updates the position, velocity, and turning rate state variables, and the LSTM-KF model updates the position and velocity state variables, which reduces the computational complexity and improves the efficiency of the algorithm.

[0124] For each model, calculate the likelihood of the Gaussian state corresponding to a given observation value. This likelihood is used to measure the degree of fit of the current observation value under each model. Combined with the model weight at the current moment, the weighted likelihood of each model is calculated. The sum of all weighted likelihoods is calculated as a normalization factor to update the model weight. Finally, the final result at the current moment is obtained by weighting the results of each model, and the above steps are repeated until the simulation ends.

[0125] The process of weighted calculation to obtain the final state estimation result is as follows:

[0126]

[0127] in, is the final state estimation result, For the j Model The estimated position at time, is the updated Markov transition probability matrix, is the covariance matrix of the j-th model.

[0128] In a specific embodiment, the time series information of the X-axis and Y-axis of the aircraft is used for verification, and the system simulation environment is Windows system and Visual Studio Code platform. The programming language is Python 3.8.

[0129] 5,000 trajectories are generated from the aircraft simulation system and constructed into a data set and a test set, and the LSTM network model is trained. The aircraft trajectory is a planar trajectory. Ensuring the accuracy of LSTM prediction is the key to improving the accuracy of the entire system.

[0130] The input data is preprocessed, each model is cycled to predict and update the Kalman filter, the model probability is updated, and the mixed multi-model results are obtained to obtain the final output.

[0131] like Figure 4 As shown in the figure, after the 50th time step, the weight of LSTM-KF is significantly greater than that of the traditional kinematic model Kalman filter, indicating that when the historical trajectory information is fully input, the state estimation performance of LSTM-KF is better than that of the traditional kinematic Kalman filter.

[0132] like Figure 5 As shown in the figure, the three figures represent the true values ​​in the x-direction, y-direction and z-direction of the position, and the comparison between the multi-model interaction filter without LSTM-KF and the multi-model interaction filter with LSTM-KF. It can be seen that the performance of the multi-model interaction filter with LSTM-KF is better than that of the multi-model interaction filter without LSTM-KF.

[0133] like Figure 4 and Figure 5 As shown, the state estimation method of the present application is significantly better than the traditional single kinematic model Kalman filter and the traditional multi-model interactive filter. The present application uses a long short-term memory Kalman filter to improve the upper limit of the accuracy of the multi-model interactive filter. When the deep neural network prediction results are biased at certain moments, it can switch to the traditional kinematics-based filter in time, thereby ensuring the system performance of the multi-model interactive filter.

[0134] Through the above-mentioned aircraft state estimation method of joint multi-model interactive filtering and DNN, on the one hand, the LSTM network model is trained according to the collected aircraft trajectory set, and the LSTM-KF filter uses the nonlinear function composed of the LSTM network model to replace the original state transfer matrix for state prediction update, and the LSTM-KF filter is connected to the interactive multi-model filtering framework as a filter module. This method has interactions between multiple different models and can adapt well to the complex motion mode of the aircraft, even when the aircraft motion mode jumps. It can still adapt well, and at the same time integrates the results of the deep neural network filter to improve the upper limit of the aircraft state estimation accuracy. On the other hand, the use of LSTM-KF filter improves the upper limit of the accuracy of the interactive multi-model filter. When the deep neural network prediction results are biased at certain moments, it can switch to the traditional kinematics-based filter in time, ensuring the system performance of the interactive multi-model filter.

[0135] It should be understood that the terms "center", "longitudinal", "lateral", "length", "width", "thickness", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", "clockwise", "counterclockwise", etc. in the above description indicate orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the embodiments of the present disclosure and simplifying the description, and do not indicate or imply that the referred device or element must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be understood as a limitation on the embodiments of the present disclosure.

[0136] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the features. In the description of the embodiments of the present disclosure, the meaning of "multiple" is two or more, unless otherwise clearly and specifically defined.

[0137] In the embodiments of the present disclosure, unless otherwise clearly specified and limited, the terms "installed", "connected", "connected", "fixed" and the like should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium, it can be the internal connection of two elements or the interaction relationship between two elements. For ordinary technicians in this field, the specific meanings of the above terms in the present disclosure can be understood according to specific circumstances.

[0138] In the embodiments of the present disclosure, unless otherwise clearly specified and limited, a first feature being "above" or "below" a second feature may include the first and second features being in direct contact, or may include the first and second features not being in direct contact but being in contact through another feature between them. Moreover, a first feature being "above", "above" and "above" a second feature includes the first feature being directly above and obliquely above the second feature, or simply indicates that the first feature is higher in level than the second feature. A first feature being "below", "below" and "below" a second feature includes the first feature being directly below and obliquely below the second feature, or simply indicates that the first feature is lower in level than the second feature.

[0139] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example" or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present disclosure. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification.

[0140] Those skilled in the art will readily appreciate other embodiments of the present disclosure after considering the specification and practicing the invention disclosed herein. This application is intended to cover any modification, use or adaptation of the present disclosure, which follows the general principles of the present disclosure and includes common knowledge or customary techniques in the art that are not disclosed in the present disclosure. The specification and examples are intended to be exemplary only, and the true scope and spirit of the present disclosure are indicated by the appended claims.

Claims

1. A method for aircraft state estimation combining multi-model interactive filtering and DNN, characterized in that: The method includes: Determine a number of kinematic models according to the state variables of the aircraft, and establish a number of corresponding traditional kinematic Kalman filters according to the number of kinematic models; wherein the kinematic models include a uniform speed model, a uniform acceleration model and a uniform speed turning model; Build and train the LSTM network model, and build the LSTM-KF filter based on the trained LSTM network model; Build an interactive multi-model filtering framework and add all traditional kinematic Kalman filters and LSTM-KF filters to the interactive multi-model filtering framework; The historical position information is input into the interactive multi-model filtering framework, all traditional kinematic Kalman filters and LSTM-KF filters are traversed for prediction updates, the model probability is calculated and updated, and the final state estimation result is obtained through weighted calculation; The steps of constructing and training the LSTM network model and constructing the LSTM-KF filter according to the trained LSTM network model include: Obtain the position information sequence of the aircraft, take the first M frames of the position information sequence as input and the M+1th frame as the true value to form a data set; Divide the data set into training set and test set according to the preset ratio; Build LSTM network model; Use the training set to train the LSTM network model, and use the test set to test the trained LSTM network model to obtain a trained LSTM network model; The nonlinear function of the trained LSTM network model is used to replace the state transfer matrix to update the state and form an LSTM-KF filter; The steps of inputting historical position information into the interactive multi-model filtering framework, traversing all traditional kinematic Kalman filters and LSTM-KF filters for prediction updates, calculating and updating model probabilities, and obtaining the final state estimation results through weighted calculation include: The Markov transition probability matrix is ​​used to realize the model interaction between the traditional kinematic Kalman filter and the LSTM-KF filter. The mixed estimate and covariance are calculated according to the state estimate and model probability of each traditional kinematic Kalman filter at the previous moment, and the mixed estimate is used as the initial state of the current cycle. The results of the hybrid estimation are input into all traditional kinematic Kalman filters and LSTM-KF filters for prediction and update steps; The predicted state obtained by each traditional kinematic Kalman filter is combined with the observed value to calculate the observation likelihood, the weighted likelihood of each traditional kinematic Kalman filter is calculated in combination with the model weight, and the sum of all weighted likelihoods is calculated as a normalization factor and the model probability is updated; Based on the model probability, the estimation results of each traditional kinematic Kalman filter are weighted and combined to obtain the final state estimation and covariance estimation; The prediction steps of the LSTM-KF filter are: in, is the time step The predicted value of the LSTM network model, is the inference function of the LSTM network model, is the time step The estimated value of the position, is the time step The estimated value of the position, is the time step The estimated value of the position, is the time step The predicted covariance matrix, is the updated Markov transition probability matrix, is for The Jacobian matrix of is the system noise matrix; The update steps of LSTM-KF filter are: in, is the Kalman gain, is the observation matrix, is the observation noise matrix, is at the time step The observed value of is the time step The final estimate of is the time step The covariance matrix of It is a unit array.

2. The method for aircraft state estimation combining multi-model interactive filtering and DNN according to claim 1 is characterized in that: The state variables include: The three-axis position, three-axis velocity, three-axis acceleration and turning rate of the aircraft; among which, The state variables of the uniform speed model are the three-axis position and three-axis velocity of the aircraft. The state variables of the uniform acceleration model are the three-axis position, three-axis velocity and three-axis acceleration of the aircraft. The state variables of the uniform turning model include the three-axis position, three-axis velocity and turning rate of the aircraft. The state variables of the LSTM network model are the three-axis position and three-axis velocity of the aircraft.

3. The method for aircraft state estimation combining multi-model interactive filtering and DNN according to claim 2 is characterized in that: The input of the LSTM network model is the historical trajectory time series information, and the output is the position prediction value at the current moment, that is: in, is the time step The predicted value of the LSTM network model, is the inference function of the LSTM network model, is the time step The estimated value of the position, is the time step The estimated value of the position, is the time step The estimated value of the position of The specific structure of the LSTM network model is: in, is the output of the forget gate at time t, is the sigmoid activation function, To act on the hidden state The weight matrix of To act on the current input The weight matrix of is the hidden state at time t-1, is the hidden state at time t, is the input of the network at each moment, is the network cell state at time t, is the bias vector of the forget gate, is the output of the input gate at time t, To act on the hidden state The weight matrix of To act on input The weight matrix of is the bias vector of the input gate, is the output of the output gate at time t, To act on the hidden state The weight matrix of To act on input The weight matrix of is the bias vector of the output gate, is the candidate cell state, is the tanh activation function, To act on the current input The weight matrix of To act on input The weight matrix of is the bias vector of the candidate cell state, is the network cell state at time t-1, Represents element-wise multiplication.

4. The method for aircraft state estimation combining multi-model interactive filtering and DNN according to claim 3 is characterized in that: The Markov transition probability matrix is: in, is the Markov transition probability matrix.

5. The method for aircraft state estimation combining multi-model interactive filtering and DNN according to claim 4 is characterized in that: The steps of calculating the observation likelihood according to the predicted state obtained by each traditional kinematic Kalman filter combined with the observation value, calculating the weighted likelihood of each traditional kinematic Kalman filter combined with the model weight, and calculating the sum of all weighted likelihoods as a normalization factor and updating the model probability include: Calculate the j The likelihood function of the model is: in, is the measurement error covariance matrix, is the measurement error; Calculate the normalization constant: in, is the credibility of the fused model; Calculate the j Update model probability of models: in, is the model probability at the k-1th moment.

6. The method for aircraft state estimation combining multi-model interactive filtering and DNN according to claim 5 is characterized in that: Based on the model probability, the estimation results of each traditional kinematic Kalman filter are weighted and combined to obtain the final state estimation and covariance estimation steps, including: in, is the final state estimation result, For the j Model The estimated position at time, is the updated Markov transition probability matrix, is the covariance matrix of the j-th model.

Images