Flight track prediction method under fault condition of airborne sensor

Abstract

The invention discloses a flight trajectory prediction method under an airborne sensor fault condition. The method comprises the following steps: S1, screening data for trajectory prediction in a public ADS-B aircraft end and ground station end data set; s2, optimizing the data in the step S1 by using a Student-t distribution variational Bayesian adaptive Kalman filtering model; s3, making an aircraft trajectory prediction spatio-temporal feature data set, wherein the feature data set comprises a training set, a verification set and a test set; s4, establishing a hybrid network prediction model, and training the hybrid network prediction model by using the feature data set in the S3; and S5, predicting a flight path by using the hybrid network prediction model. According to the method, the prediction model extracted by the multi-path space-time mixed features is utilized, the influence of the multi-dimensional state information of the target on target trajectory prediction is fully considered, the internal relation between the time sequence information of the target state can be reflected, and the reliability and the accuracy of model prediction can be guaranteed.

Description

Technical Field

[0001] The present invention relates to the technical field of trajectory prediction, and in particular to a flight trajectory prediction method in the event of an airborne sensor failure. Background Art

[0002] In recent years, Automatic Dependent Surveillance-Broadcast (ADS-B) technology has flourished and has become a promising technology in the next generation of air traffic control. Compared with traditional radar, the ADS-B system does not require manual operation or query, and has the advantages of high positioning accuracy, high data update rate, and low maintenance cost. However, since ADS-B data comes from the GPS system rather than the key navigation information such as position, altitude, and speed calculated autonomously by the aircraft flight control system, and GPS is easily subject to intentional or unintentional illegal interference, once the onboard sensor failure causes the aircraft end data to fail, it will greatly affect aviation safety. Therefore, reliable short-term and long-term prediction of aircraft trajectories can improve flight safety in busy airspace.

[0003] Research and development of flight trajectory prediction methods can provide decision support for controllers and avoid accidents safely and efficiently. At present, advanced flight trajectory prediction methods are mainly divided into two types: prediction based on physical models and prediction based on data-driven models. Among them, modeling based on physical models requires clear input of modeling flight performance and real-time aircraft status, which has become an obstacle to modeling based on physical models. In recent years, with the rapid development of artificial intelligence, scholars have gradually begun to use deep neural network models to learn and predict flight trajectories, and have also achieved high prediction accuracy. However, previous studies often did not consider them comprehensively. Either only a single neural network model was used to predict flight trajectories, and the prediction accuracy was relatively low; or a combined neural network was used, but the impact of the target's multi-dimensional state information on the target trajectory prediction was not considered, and the intrinsic connection between the time series information reflecting the target state was not discovered. Summary of the invention

[0004] The purpose of the present invention is to provide a flight trajectory prediction method in the event of an airborne sensor failure, which solves the above-mentioned technical problems and makes the prediction result more accurate.

[0005] To achieve the above object, the present invention provides the following technical solution: a method for predicting a flight trajectory in the event of an airborne sensor failure, comprising the following steps:

[0006] S1. Filter out the data used for trajectory prediction from the public ADS-B aircraft and ground station data sets;

[0007] S2, optimize the data in S1 using the Student-t distribution variational Bayesian adaptive Kalman filter model;

[0008] S3, creating a spatiotemporal feature dataset for aircraft trajectory prediction, wherein the feature dataset includes a training set, a validation set, and a test set;

[0009] S4, establish an N-Inception-LSTM-ATT hybrid network prediction model, and use the feature data set in S3 to train the N-Inception-LSTM-ATT hybrid network prediction model;

[0010] S5. Use the N-Inception-LSTM-ATT hybrid network prediction model to predict the flight trajectory.

[0011] In some embodiments, S1 comprises the following steps:

[0012] S11. Process the internationally public ADS-B aircraft data set, visualize the aircraft flight trajectory, manually judge the validity of the trajectory, eliminate invalid and discontinuous trajectories, and screen out continuous and complete data;

[0013] S12. Visually analyze the ADS-B ground station data set from the perspectives of the receiver clock and received signal strength, identify and remove receiver data with clock damage and received signal strength failure, and retain other data;

[0014] S13, integrate the final aircraft-side data set and ground station-side data set in S12 into a new data set to ensure the reliability of the data and facilitate the production of subsequent training sets, verification sets, and test sets.

[0015] In some embodiments, S12 includes the following steps:

[0016] First, we filter the data from the receiver clock aspect, draw a two-dimensional graph of the server timestamp and the receiver timestamp in Matlab, and filter out the data on the straight line with a slope of 1;

[0017] Secondly, filter the data from the perspective of received signal strength and draw a graph of the relationship between the receiving distance and the received signal strength. Theoretically, the closer the distance, the stronger the received signal strength. The received signal strength is linearly related to the logarithm of the receiving distance with a base of 10. Filter out the data that meets this rule.

[0018] Finally, the receiver data that does not meet the above conditions is eliminated, and the other data is finally retained.

[0019] In some embodiments, in S2, the state and measurement noise parameters can be recursively solved by using a Bayesian adaptive Kalman filter model, effectively solving the unknown time-varying problem of the original data measurement noise, and further optimizing the data quality. The Bayesian adaptive Kalman filter model is:

[0020]

[0021] α k,i =α k-1,i +0.5;

[0022] v k,i =α k,i / β k,i ;

[0023]

[0024] D k =(Z k -H k X k )(Z k -H k X k ) T ;

[0025]

[0026] Taking longitude as an example, where P k-1 is the k-1 moment posterior estimated state covariance matrix; ρ is the attenuation coefficient, ρ∈(0,1]; Λ is the precision matrix; X k is the estimated value of the posterior state of longitude at the kth moment; is the prior state estimate of the longitude at the kth moment; F k is the state transfer matrix; X k-1 is the posterior state estimate of the longitude at the k-1th moment; is the transposed result of the state transfer matrix; Q k is the process noise variance matrix; is the prior estimated state covariance matrix at time k; is the prior value of the hyperparameter α of the Student-t distribution in dimension i at time k; α k-1,i is the posterior value of the hyperparameter α of the Student-t distribution in dimension i at time k-1; is the hyperparameter β prior value of the Student-t distribution in dimension i at time k; β k-1,i is the posterior value of the hyperparameter β of the Student-t distribution in dimension i at time k-1; α k,i is the posterior value of the hyperparameter α of the Student-t distribution in dimension i at time k; v k,i is the degree of freedom of dimension i at time k; βk,i is the posterior value of the hyperparameter β of the Student-t distribution in dimension i at time k; M k An auxiliary matrix for convenient calculation; H k is the measurement matrix; is the transposed result of the measurement matrix; K k is the filter gain matrix; is the inverse matrix of the auxiliary matrix; Z k is the observation vector; is the transpose of the filter gain matrix; D k is the posterior estimated covariance matrix at time k; is the k-time trace matrix; k is the auxiliary random variable introduced into the measurement distribution at time k; d is the dimension of the observation vector; is the value of the i-th dimension of the trace matrix at time k.

[0027] In some embodiments, S3 includes the following steps:

[0028] S31, normalizing the data set obtained in S2 so that the normalized data is between the interval [0, 1];

[0029] S32, selecting the features of a single time node; the input features of the single time node are the time, receiver signal strength, longitude, latitude and altitude of the ground station; the output features are the longitude, latitude and altitude of the aircraft;

[0030] S33, selecting a suitable sliding window size;

[0031] S34. Shuffle the data set to finally obtain appropriate training set, validation set and test set.

[0032] In some embodiments, the N-Inception-LSTM-ATT hybrid neural network fully considers the impact of the target multi-dimensional state information on the target trajectory prediction, and can also reflect the intrinsic connection between the time series information of the target state. After training, more accurate and reliable prediction results can be obtained. S4 includes the following steps:

[0033] S41, vertically splice the convolutional neural network Inception, the long short-term memory neural network LSTM and the self-attention mechanism submodules to construct the N-Inception-LSTM-ATT hybrid neural network, where the optimizer is Adam and the loss function is MSE;

[0034] S42, set the number of iterations, train the N-Inception-LSTM-ATT hybrid neural network to minimize the loss function MSE;

[0035] S43. Adjust the network parameters to obtain the N-Inception-LSTM-ATT hybrid neural network prediction model.

[0036] In some embodiments, S41 includes the following steps:

[0037] First, the training set is input into the convolutional neural network Inception to extract the spatial features of the target state information, and the results are input into the long short-term memory neural network LSTM, which has stronger processing capabilities for long time series, to mine the temporal features of the data; then, the ATT module is connected before the output layer to fully consider all input vectors, capture the dependencies in the sequence data, and improve the performance of the model in the sequence data; the self-attention mechanism uses the inherent information in the data features to interact with attention, and calculates the query vector Q (Query), key vector K (Key) and value vector V (Value) through multiple inputs. The calculation expression is:

[0038]

[0039] Where, d k is the dimension of K, and its purpose is to scale the value after the dot product of the query vector Q and the key vector K.

[0040] Compared with the prior art, the present invention has the following beneficial effects:

[0041] In the case where the aircraft-side data is unavailable due to an abnormality of the airborne sensor, the present invention uses the signal strength of the ground receiver to provide a certain degree of reliable prediction value, thereby reducing time and labor costs;

[0042] The present invention utilizes a prediction model for multi-channel spatiotemporal hybrid feature extraction, fully considers the impact of target multidimensional state information on target trajectory prediction, and can also reflect the intrinsic connection between the time series information of the target state, thereby ensuring the reliability and accuracy of the model prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 It is a schematic diagram of the process of the present invention;

[0044] Figure 2 Schematic diagram of the process of Example S1 of the present invention;

[0045] Figure 3 This is a schematic diagram of the process of Example S3 of the present invention;

[0046] Figure 4 Schematic diagram of the process of S32 of the embodiment of the present invention;

[0047] Figure 5 Schematic diagram of the process of S33 of the embodiment of the present invention;

[0048] Figure 6 It is a schematic diagram of the structure of the N-Inception-LSTM-ATT hybrid neural network model of the present invention;

[0049] Figure 7 A schematic diagram of a process of trajectory prediction according to an embodiment of the present invention;

[0050] Figure 8 shows the longitude and latitude prediction results of different models in an embodiment of the present invention; Figure 8(a) is a comparison chart of the three-dimensional prediction results of the 2-Inception and 2-Inception-LSTM models; Figure 8(b) is a comparison chart of the three-dimensional prediction results of Inception-LSTM and 2-Inception-LSTM; Figure 8(c) is a comparison chart of the three-dimensional prediction results of 2-Inception-LSTM and 2-Inception-LSTM-ATT. DETAILED DESCRIPTION

[0051] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0052] See also Figure 1 , a flight trajectory prediction method in the case of airborne sensor failure, comprising the following steps:

[0053] S1. Filter out data for trajectory prediction from the public ADS-B aircraft and ground station data sets.

[0054] like Figure 2 As shown, the specific steps include:

[0055] S11. Process the internationally disclosed ADS-B aircraft data set, import the aircraft longitude, latitude and altitude data into Matlab software, use the point drawing method to draw a three-dimensional trajectory to visualize the aircraft flight trajectory, manually remove the trajectory with interrupted signals and obviously erroneous trajectories, so as to judge the validity of the trajectory, remove invalid and intermittent trajectories, and screen out continuous, complete and highly reliable data.

[0056] S12. Visually analyze the ADS-B ground station data set from the perspectives of the receiver clock and received signal strength, identify and eliminate receiver data with clock damage and received signal strength failure, and retain high-reliability data.

[0057] Specifically, first, filter the data from the perspective of the receiver clock, and draw a two-dimensional graph of the server timestamp and the receiver timestamp in Matlab, with the server timestamp on the horizontal axis and the receiver timestamp on the vertical axis. Theoretically, the graph that should be presented is a straight line with a slope of 1. As long as the data does not deviate much, it can be retained; then filter the data from the perspective of received signal strength, and draw a two-dimensional graph of the receiving distance and the received signal strength, with the horizontal axis being the receiving distance and the vertical axis being the receiving strength. In theory, the received signal strength is affected by the receiving distance and the propagation environment. When the receiving distance becomes smaller, the receiving strength becomes larger accordingly. When the receiving distance becomes smaller, the data with unchanged or increased received signal strength needs to be eliminated. The received signal strength is linearly related to the logarithm of the receiving distance with a base of 10. According to the above theory, the receiver data that does not meet the above conditions is filtered out and eliminated, and finally the highly reliable data is retained.

[0058] S13, integrate the final aircraft-side data set and ground station-side data set in S12 into a new data set to ensure data reliability and facilitate the production of subsequent training sets, verification sets, and test sets.

[0059] S2, optimize the data in S1 using the Student-t distribution variational Bayesian adaptive Kalman filter model;

[0060] Since the original data set will be affected by the characteristics of the sensor itself and the complex environment around it, it is necessary to use the Student-t distribution variational Bayesian adaptive Kalman filter model to optimize it and solve the problem of unknown time-varying noise and the existence of outliers.

[0061] Longitude, latitude and altitude are modeled separately, where the state variable and observation vector of longitude can be selected as the longitude values ​​of three different sensors, and the latitude and altitude are modeled in the same way as above. The Student-t distribution variational Bayes adaptive Kalman filter model is mainly divided into two aspects: prediction and update. By changing the values ​​of α and β to change v and λ, the purpose of adaptive estimation is finally achieved. The prediction and update formulas are as follows:

[0062]

[0063] α k,i =α k-1,i +0.5;

[0064] v k,i =α k,i / β k,i ;

[0065]

[0066] D k =(Z k -H kX k )(Z k -H k X k ) T ;

[0067]

[0068] Taking longitude as an example, where P k-1 is the k-1 moment posterior estimated state covariance matrix; ρ is the attenuation coefficient, ρ∈(0,1]; Λ is the precision matrix; X k is the estimated value of the posterior state of longitude at the kth moment; is the prior state estimate of the longitude at the kth moment; F k is the state transfer matrix; X k-1 is the posterior state estimate of the longitude at the k-1th moment; is the transposed result of the state transfer matrix; Q k is the process noise variance matrix; is the prior estimated state covariance matrix at time k; is the prior value of the hyperparameter α of the Student-t distribution in dimension i at time k; α k-1,i is the posterior value of the hyperparameter α of the Student-t distribution in dimension i at time k-1; is the hyperparameter β prior value of the Student-t distribution in dimension i at time k; β k-1,i is the posterior value of the hyperparameter β of the Student-t distribution in dimension i at time k-1; α k,i is the posterior value of the hyperparameter α of the Student-t distribution in dimension i at time k; v k,i is the degree of freedom of dimension i at time k; β k,i is the posterior value of the hyperparameter β of the Student-t distribution in dimension i at time k; M k An auxiliary matrix for convenient calculation; H k is the measurement matrix; is the transposed result of the measurement matrix; K k is the filter gain matrix; is the inverse matrix of the auxiliary matrix; Z k is the observation vector; is the transpose of the filter gain matrix; D k is the posterior estimated covariance matrix at time k; is the k-time trace matrix; k is the auxiliary random variable introduced into the measurement distribution at time k; d is the dimension of the observation vector; is the value of the i-th dimension of the trace matrix at time k.

[0069] S3, creating a space-time feature dataset for aircraft trajectory prediction, wherein the feature dataset includes a training set, a validation set, and a test set. Figure 3 As shown, the following steps are included:

[0070] S31, normalize the data set obtained in S2 so that the normalized data is between [0, 1]; the normalization process is to normalize the maximum and minimum values ​​of each field in the data set separately:

[0071]

[0072] In the formula, X is the data of a field, min is the minimum value in the field, is the maximum value, and N is the normalized result.

[0073] S32, select the features of a single time node; Figure 4 As shown, the input features of the single time node select 3 ground station data, each of which includes time, receiver signal strength, longitude, latitude and altitude, a total of 15 input features; the output features are the longitude, latitude and altitude of the aircraft, a total of 3 output features.

[0074] S33, select a suitable sliding window size; Figure 5 As shown in the figure, in order to make full use of the time dimension characteristics of the data set, the time window is set to 64. Combined with the input features selected in the previous step, the slider is 64×15.

[0075] S34. Shuffle the data set to finally obtain appropriate training set, validation set and test set.

[0076] S4, establish the N-Inception-LSTM-ATT hybrid network prediction model, and use the feature data set in S3 to train the N-Inception-LSTM-ATT hybrid network prediction model. Figure 6 As shown, the following steps are included:

[0077] S41. Vertically connect the convolutional neural network Inception, the long short-term memory neural network LSTM and the self-attention mechanism sub-modules to construct the N-Inception-LSTM-ATT hybrid neural network, where the optimizer is Adam and the loss function is MSE.

[0078] First, the training set is input into the convolutional neural network Inception to extract the spatial features of the target state information, and the obtained results are input into the long short-term memory neural network LSTM, which has stronger processing capabilities for long time series, to mine the temporal features of the data; then, the ATT module is connected before the output layer to fully consider all input vectors, capture the dependencies in the sequence data, and improve the performance of the model in the sequence data; the long short-term memory neural network LSTM consists of a total of 128 memory units, followed by the Dropout layer with a parameter of 0.5 (indicating that 50% of the neurons are discarded). The Dropout layer reduces the dependencies between neurons by randomly discarding a part of the neurons during the training process, thereby reducing the overfitting of the model to the training data and forcing the neural network to learn more robust features, thereby improving the generalization ability of the model.

[0079] The self-attention mechanism uses the inherent information in the data features to perform attention interaction, and calculates the query vector Q (Query), key vector K (Key) and value vector V (Value) through multiple inputs. The calculation expression is:

[0080]

[0081] Where, d k is the dimension of K, and its purpose is to scale the value after the dot product of the query vector Q and the key vector K.

[0082] S42, set the number of iterations, train the N-Inception-LSTM-ATT hybrid neural network to minimize the loss function MSE; Figure 7 As shown, the following steps are included:

[0083] Input the training set data into the N-Inception-LSTM-ATT hybrid neural network, observe the Loss value of the training set and the Loss value of the test set, and save it after judging whether the hybrid neural network is fitted and meets the accuracy. When judging whether the hybrid neural network is fitted, the optimizer is Adam and the loss function is MSE. The loss function MSE is used as the evaluation indicator, and the formula is as follows:

[0084]

[0085] In the formula, y i is the true value of the ith data sample, is the corresponding predicted value, n is the number of samples, the smaller the value of the loss function MSE, the closer the model's prediction result is to the true value, and the better the model's performance is. Set an appropriate number of iterations and train the network to minimize the loss function MSE, and the loss curve shows a normal downward trend.

[0086] S43. Adjust the network parameters to obtain the N-Inception-LSTM-ATT hybrid neural network prediction model.

[0087] When judging whether the model meets the accuracy, it is necessary to adjust the network parameters, such as the learning rate lr, batch size batch_size and number of iterations, where lr is set to 0.001, batch_size is set to 4000, and the number of training rounds is set to 200. Finally, the N-Inception-LSTM-ATT hybrid neural network prediction model with the highest accuracy is obtained.

[0088] S5. Use the N-Inception-LSTM-ATT hybrid network prediction model to predict the flight trajectory.

[0089] In a specific embodiment, the present application is compared with the model in the prior art as follows:

[0090] Table 1 Experimental group parameter settings

[0091]

[0092] As shown in Figure 8, Figure 8(a) shows the comparison of the three-dimensional prediction results of the 2-Inception and 2-Inception-LSTM models; Figure 8(b) shows the comparison of the three-dimensional prediction results of Inception-LSTM and 2-Inception-LSTM; Figure 8(c) shows the comparison of the three-dimensional prediction results of 2-Inception-LSTM and 2-Inception-LSTM-ATT. By comparing multiple different models, we can intuitively see that the prediction accuracy of the N-Inception-LSTM-ATT model is higher and closer to the real trajectory.

[0093] Taking into account the situation where the data on the aircraft side is unavailable due to a malfunction of the onboard sensor, features such as the receiver signal strength at the ground receiving end are selected for prediction. The N-Inception-LSTM-ATT model can provide reliable and accurate prediction values ​​within a certain range, reducing time and labor costs.

Claims

1. A method for predicting flight trajectory in the event of an airborne sensor failure, characterized in that: The following steps are involved: S1. Filter out the data used for trajectory prediction from the public ADS-B aircraft and ground station data sets; S2, optimize the data in S1 using the Student-t distribution variational Bayesian adaptive Kalman filter model; S3, creating a spatiotemporal feature dataset for aircraft trajectory prediction, wherein the feature dataset includes a training set, a validation set, and a test set; S4, establish an N-Inception-LSTM-ATT hybrid network prediction model, and use the feature data set in S3 to train the N-Inception-LSTM-ATT hybrid network prediction model; S5. Use the N-Inception-LSTM-ATT hybrid network prediction model to predict the flight trajectory.

2. The method for predicting flight trajectory in the event of an airborne sensor failure according to claim 1, characterized in that: S1 includes the following steps: S11. Process the internationally public ADS-B aircraft data set, visualize the aircraft flight trajectory, manually judge the validity of the trajectory, eliminate invalid and discontinuous trajectories, and screen out continuous and complete data; S12. Visually analyze the ADS-B ground station data set from the perspectives of the receiver clock and received signal strength, identify and remove receiver data with clock damage and received signal strength failure, and retain other data; S13, integrate the final aircraft-side data set and ground station-side data set in S12 into a new data set to ensure the reliability of the data and facilitate the production of subsequent training sets, verification sets, and test sets.

3. The flight trajectory prediction method in the case of an airborne sensor failure according to claim 2, characterized in that: S12 includes the following steps: First, we filter the data from the receiver clock aspect, draw a two-dimensional graph of the server timestamp and the receiver timestamp in Matlab, and filter out the data on the straight line with a slope of 1; Secondly, filter the data in terms of received signal strength and plot the relationship between receiving distance and received signal strength; Finally, the receiver data that does not meet the above conditions is eliminated, and the other data is finally retained.

4. The method for predicting flight trajectory in the event of an airborne sensor failure according to claim 1, characterized in that: In S2, the Bayesian adaptive Kalman filter model is: α k,i ＝α k-1,i +0.5； v k,i ＝α k,i / β k,i ； D k =(Z k -H k X k )(Z k -H k X k ) T ; Taking longitude as an example, where P k-1 is the posterior estimated state covariance matrix at time k-1; ρ is the attenuation coefficient, ρ∈(0,1]; Λ is the precision matrix; X k is the estimated value of the posterior state of longitude at the kth moment; is the prior state estimate of the longitude at the kth moment; F k is the state transfer matrix; X k-1 is the posterior state estimate of the longitude at the k-1th moment; is the transposed result of the state transfer matrix; Q k is the process noise variance matrix; is the prior estimated state covariance matrix at time k; is the prior value of the hyperparameter α of the Student-t distribution in dimension i at time k; α k-1,i is the posterior value of the hyperparameter α of the Student-t distribution in dimension i at time k-1; is the hyperparameter β prior value of the Student-t distribution in dimension i at time k; β k-1,i is the posterior value of the hyperparameter β of the Student-t distribution in dimension i at time k-1; α k,i is the posterior value of the hyperparameter α of the Student-t distribution in dimension i at time k; v k,i is the degree of freedom of dimension i at time k; β k,i is the posterior value of the hyperparameter β of the Student-t distribution in dimension i at time k; M k An auxiliary matrix for convenient calculation; H k is the measurement matrix; is the transposed result of the measurement matrix; K k is the filter gain matrix; is the inverse matrix of the auxiliary matrix; Z k is the observation vector; is the transpose of the filter gain matrix; D k is the posterior estimated covariance matrix at time k; is the k-time trace matrix; k is the auxiliary random variable introduced into the measurement distribution at time k; d is the dimension of the observation vector; is the value of the i-th dimension of the trace matrix at time k.

5. The method for predicting flight trajectory in the event of an airborne sensor failure according to claim 1, characterized in that: S3 includes the following steps: S31, normalizing the data set obtained in S2 so that the normalized data is between the interval [0, 1]; S32, selecting features of a single time node; The input features of the single time node are selected from the time, receiver signal strength, longitude, latitude and altitude of the ground station; The output features are the longitude, latitude, and altitude of the aircraft; S33, selecting a suitable sliding window size; S34. Shuffle the data set to finally obtain the training set, validation set and test set.

6. The method for predicting flight trajectory in the event of an airborne sensor failure according to claim 1, characterized in that: S4 includes the following steps: S41, vertically splice the convolutional neural network Inception, the long short-term memory neural network LSTM and the self-attention mechanism submodules to construct the N-Inception-LSTM-ATT hybrid neural network, where the optimizer is Adam and the loss function is MSE; S42, set the number of iterations, train the N-Inception-LSTM-ATT hybrid neural network to minimize the loss function MSE; S43. Adjust the network parameters to obtain the N-Inception-LSTM-ATT hybrid neural network prediction model.

7. The method for predicting flight trajectory in the event of an airborne sensor failure according to claim 6, characterized in that: S41 includes the following steps: First, the training set is input into the convolutional neural network Inception to extract the spatial features of the target state information, and the obtained results are input into the long short-term memory neural network LSTM to mine the temporal features of the data; Then, the ATT module is connected before the output layer to fully consider all input vectors, capture the dependencies in the sequence data, and improve the performance of the model in the sequence data; the self-attention mechanism uses the inherent information in the data features to interact with attention, and calculates the query vector Q (Query), key vector K (Key) and value vector V (Value) through multiple inputs. The calculation expression is: Where, d k is the dimension of K, and its purpose is to scale the value after the dot product of the query vector Q and the key vector K.

Images