A real-time UAV trajectory ensemble prediction method based on transfer learning

Abstract

The present invention discloses a method for integrated prediction of real-time trajectory of unmanned aerial vehicles based on transfer learning, including: collecting trajectory data of various types of light rotor unmanned aerial vehicles; processing the trajectory data of each type of unmanned aerial vehicle into a time series slice data set; dividing the sample set into a time series slice data set of unmanned aerial vehicle trajectory of the predicted target and a time series slice data set of unmanned aerial vehicle trajectory other than the predicted target; performing dimension reduction of trajectory slices; screening similar trajectories of training samples; obtaining a D-GRU pre-trained trajectory prediction model based on sample screening; obtaining an integrated prediction model; performing real-time prediction of the predicted target trajectory slices based on the integrated prediction model; and evaluating the unmanned aerial vehicle trajectory prediction results. The method of the present invention improves the accuracy of trajectory prediction based on data-driven for target unmanned aerial vehicles lacking historical trajectory information, lays a foundation for tasks such as flight conflict detection and dangerous behavior identification, and provides effective decision support for urban low-altitude airspace management.

Description

Technical Field

[0001] The present invention belongs to the technical field of unmanned aerial vehicle (UAV) traffic management, and in particular relates to a UAV real-time trajectory integrated prediction method based on transfer learning. Background Art

[0002] With the gradual opening of the country's low-altitude airspace, drones are rapidly popularizing in various industries. In order to achieve safe, efficient and orderly operation in urban low-altitude airspace, it is necessary to improve the technical means of drone safety management. In mission scenarios that require real-time trajectory prediction, such as urban low-altitude safety management and airport no-fly zone safety control, it is difficult to obtain rich target information in a concise and efficient manner when facing non-cooperative targets that appear for the first time. If we can achieve the accuracy advantage compared to the kinematic modeling method through transfer learning of different drone trajectories and meet the requirements of real-time prediction, it will lay the foundation for tasks such as drone flight conflict detection and dangerous behavior identification, and provide effective decision-making support for tasks such as urban low-altitude airspace management and safety control of special areas such as airports.

[0003] When facing tasks where there are not a large number of real trajectories of the predicted target, the principles of methods such as particle filtering ensure that they have good prediction effects on uniform speed and uniform acceleration flight trajectories, but their prediction accuracy is greatly reduced in scenarios with complex and changeable motion behaviors; and the trajectory prediction model based on data mining and machine learning has stronger nonlinear expression capabilities than the method based on the motion equation combined with particle filtering, and has outstanding prediction effects in tasks such as complex and changeable UAV motion. In actual UAV traffic management tasks, when facing prediction targets that do not have sufficient real trajectory samples, how to improve the prediction accuracy and timeliness of data-driven models through transfer learning and integrated prediction, and enhance their applicability in real-time trajectory prediction tasks, is one of the important issues that need to be solved from theory to practical application of data-driven trajectory prediction methods. Summary of the invention

[0004] In view of the above-mentioned deficiencies in the prior art, the purpose of the present invention is to provide a real-time integrated trajectory prediction method for drones based on transfer learning, so as to solve the problem of decreased prediction accuracy of the data-driven trajectory prediction model in the prior art in tasks with insufficient historical trajectory samples; the method of the present invention improves the efficiency of drone safety management in urban low-altitude airspace.

[0005] To achieve the above object, the technical solution adopted by the present invention is as follows:

[0006] The present invention provides a method for real-time trajectory integration prediction of unmanned aerial vehicles based on transfer learning, and the steps are as follows:

[0007] (1) Collect trajectory data of various types of light rotary-wing UAVs and calculate the performance parameters of each type of UAV;

[0008] (2) Processing the trajectory data of each type of drone collected in the above step (1) into a time series slice data set;

[0009] (3) According to the performance parameter constraints of each type of UAV, the time series slice data set of each type of UAV trajectory obtained in step (2) is updated, and the sample set is divided into the time series slice data set of the UAV trajectory of the prediction target (sample set 1) and the time series slice data set of the UAV trajectory other than the prediction target (sample set 2);

[0010] (4) For the UAV trajectory time series slice dataset of the predicted target and the UAV trajectory time series slice dataset other than the predicted target, trajectory slice dimension reduction based on statistical measurement is performed respectively;

[0011] (5) Screening similar trajectories of training samples for the time series sliced ​​dataset of drone trajectories except for the predicted target after dimensionality reduction;

[0012] (6) Using the D-GRU trajectory prediction model to perform model pre-training, and obtaining a D-GRU pre-trained trajectory prediction model based on sample screening;

[0013] (7) Integrate the D-GRU pre-trained trajectory prediction model based on sample screening with the trajectory prediction model based on particle filtering to obtain an integrated prediction model;

[0014] (8) Real-time prediction of target trajectory slices based on the integrated prediction model;

[0015] (9) Evaluate the UAV trajectory prediction results.

[0016] Furthermore, each track data in step (1) includes multiple track points, and the information recorded at each track point includes: world coordinated timestamp, drone identity ID, longitude, latitude, flight altitude, and flight ground speed.

[0017] Furthermore, the performance parameters in step (1) include: net weight, maximum static wind speed, maximum wind speed, maximum climb rate and maximum descent rate.

[0018] Furthermore, the specific process of step (2) is as follows:

[0019] (21) Generate equally spaced time series data;

[0020] The trajectory data of each type of UAV collected in step (1) are processed into sequence data with equal time intervals through linear interpolation;

[0021] (22) Generate a time series slice dataset;

[0022] A sliding time window is used to divide the sequence data of each equal time interval obtained in step (21) into time series slices of fixed length, thereby generating a time series slice dataset that contains both input and output data in the trajectory prediction model.

[0023] Furthermore, the specific process of step (3) is as follows:

[0024] (31) Determine speed constraints;

[0025] Solve the vertical component of the estimated velocity of each trajectory point in the time series slice data set obtained in step (2), with the vertical component satisfying the maximum climb / maximum descent rate in the performance parameters of each model of drone as a constraint;

[0026] The estimated velocity V of the i-th trajectory point i It is expressed as:

[0027]

[0028] Where Δt represents the trajectory sampling interval, P i+1 and P i-1 The relative coordinates of the i+1th and i-1th trajectory points relative to the first trajectory point, respectively. The first trajectory point of each trajectory time series slice is set as the coordinate origin on the corresponding slice;

[0029] (32) Filtering time series slice datasets;

[0030] According to the flight speed limit (i.e., the maximum climb / maximum descent rate constraint) of each type of UAV in step (31), the time series slice data set that meets the flight speed limit in step (2) is selected;

[0031] (33) Two different types of slice sample sets are divided as follows:

[0032] (331) Sample set 1: The time series slice dataset of the drone trajectory of the predicted target in step (32);

[0033] (332) Sample set 2: Sample set 2 is obtained by mixing the drone trajectory time series slice dataset except the predicted target model in step (32).

[0034] Furthermore, the specific process of step (4) is as follows:

[0035] (41) Setting the coordinate origin of the trajectory slice;

[0036] The coordinates of the first track point of each track time series slice after division are marked as (0,0,0), that is, set as the coordinate origin on the corresponding slice;

[0037] (42) Based on feature statistics, sample set 1 and sample set 2 are reduced in dimension respectively;

[0038] The speed and heading change of each trajectory point in each trajectory slice relative to the coordinate origin in step (41) are calculated respectively, and then the flight speed variance V of each trajectory slice is calculated. var and the accumulated heading change A chg :

[0039]

[0040]

[0041] Where, T in is the time step of the input information, V i is the velocity of the ith trajectory point, V s is the distance from the 2nd point to the Tth point in the trajectory slice in -1 point average flight speed, A i is the heading variable of the i-th trajectory point, and the flight speed variance V var and the accumulated heading change A chg As a potential feature that can comprehensively express the time series slices in low-dimensional space, the time series slices in step (33) are expressed in low-dimensional space.

[0042] Furthermore, the specific process of step (5) is as follows:

[0043] (51) Quantile division;

[0044] For the potential feature flight speed variance V var and the accumulated heading change A chg , starting from the 0 quantile, m and n equally spaced quantiles are set separately, and the UAV mixed trajectory time series slice dataset except the predicted target after dimensionality reduction is divided into m×n subsets;

[0045] (52) Solve the sample coordinates of time series slices;

[0046] For any time series slice sample in the mixed sample set in step (51), based on the flight speed variance and cumulative heading change of the sample, the potential characteristic flight speed variance V var and the accumulated heading change A chg The position of the quantile point, and the coordinates of the sample are obtained;

[0047] (53) Solve the coordinate matrix of the mixed sample set;

[0048] Combined with step (52), the coordinates of all time series slice samples in the mixed sample set in step (51) are obtained, and the coordinate matrix of the mixed trajectory time series slice data set of the drone except the predicted target after dimensionality reduction is obtained, and then divided into m×n subsets, among which the subset with large values ​​of m and n will contain the small subset, based on the flight speed variance V var and the accumulated heading change A chg The numerical similarity measures the distance between training samples and performs subset screening on the UAV mixed trajectory time series slice dataset except for the predicted target after dimensionality reduction.

[0049] Furthermore, the specific process of step (6) is as follows:

[0050] (61) Setting the D-GRU trajectory prediction model parameters;

[0051] (611) Setting model input layer parameters;

[0052] Determine the number of training samples in a single batch and the input data dimension of each sample; initialize the input layer weights and biases;

[0053] (612) Setting model network layer parameters;

[0054] Determine the number of hidden layers, the number of neurons in a single layer, and the time step of input information; initialize the network layer weights and biases; initially select elu as the activation function;

[0055] (613) Setting model output layer parameters;

[0056] Determine the output data dimension; initialize the output layer weights and biases; initially select elu as the activation function;

[0057] (62) Construct D-GRU trajectory prediction model training samples;

[0058] The sample set 2 in step (332) is divided into a classification set S and an integration set U in a ratio of 7:3; according to steps (4) and (53), S and U are respectively reduced in dimension and the coordinate matrix is ​​solved, and then S is divided into a set S containing m×n sample subsets mn ;

[0059] (63) Setting the loss function of the D-GRU trajectory prediction model;

[0060] The L2 regularization method is used as the loss function of mean square error (MSE);

[0061] (64) Setting the D-GRU trajectory prediction model optimizer;

[0062] Select Adam as the model optimizer with a learning rate of 1.0e-4;

[0063] (65) Train and save the D-GRU trajectory prediction model;

[0064] Take the set S in step (62) mn The input samples are constructed from each subset of , the D-GRU trajectory prediction model is trained and saved, and the D-GRU pre-trained trajectory prediction model based on sample screening is obtained (model C).

[0065] Furthermore, the specific process of step (7) is as follows:

[0066] (71) Solve for each predicted value of the integrated set U;

[0067] For each sample u in the integrated set U in step (62), the prediction value p is obtained by combining the D-GRU pre-trained trajectory prediction model (model C) based on sample screening in step (65). u,0 , the trajectory prediction model based on particle filtering (model D) obtains the predicted value p u,1 , and get the predicted value p of the integrated set U u,0 and p u,1 ;

[0068] (72) Integrated prediction model training and storage;

[0069] The two predicted values ​​of the integrated set U in step (71) and its own input data are used as the input of the integrated prediction model, and the true output of the integrated set U is used as the output of the integrated prediction model. The integrated weights of the D-GRU pre-trained trajectory prediction model (model C) based on sample screening and the trajectory prediction model (model D) based on particle filtering in step (71) are trained and saved, and the UAV trajectory integrated prediction model (model E) based on data migration is obtained.

[0070] Furthermore, the specific process of step (8) is as follows:

[0071] (81) Loading the pre-trained model;

[0072] Load the D-GRU pre-trained trajectory prediction model (model C) based on sample screening in step (65) and the drone trajectory integration prediction model (model E) based on data migration in step (72); obtain real-time trajectory slice s;

[0073] (82) Real-time trajectory slicing dimensionality reduction;

[0074] According to the feature expression method in step (4), s is reduced to a vector (V var,s , A chg,s );

[0075] (83) Solve the prediction values ​​of the trajectory prediction model based on sample screening D-GRU pre-training (model C) and the trajectory prediction model based on particle filtering (model D) respectively;

[0076] The predicted value p is obtained according to the D-GRU pre-trained trajectory prediction model (model C) based on sample screening loaded in step (81) s,0 , the trajectory prediction model based on particle filtering (model D) obtains the predicted value p s,1 ;

[0077] (84) solving the prediction value of the integrated prediction model;

[0078] Based on the data migration-based UAV trajectory integrated prediction model (model E) loaded in step (81), the final prediction value p of the trajectory slice s is obtained s .

[0079] Furthermore, the specific process of step (9) is as follows:

[0080] (91) Calculate trajectory prediction accuracy;

[0081] According to the final predicted values ​​of each type of drone obtained in step (84), the average Euclidean error E between the predicted coordinates and the true coordinates in three-dimensional space is calculated. e , which is used as an evaluation index to evaluate the accuracy of UAV trajectory prediction;

[0082] (92) Comparison of prediction accuracy and time consumption measurement;

[0083] (921) Comparison of the accuracy of different prediction models;

[0084] Model A: D-GRU trajectory prediction model based on predicted target trajectory data, as follows:

[0085] For the sample set 1 in the step (331) after dimensionality reduction, i.e., the time series slice data set for predicting the target trajectory, the prediction result is obtained using the trajectory prediction model based on D-GRU;

[0086] Model B: Trajectory prediction model based on D-GRU pre-training without sample screening, as follows:

[0087] When constructing the training samples of the D-GRU trajectory prediction model in step (62), no sample screening is performed, that is, m=n=1 is set when the classification set S is divided. According to step (6), the trajectory prediction model based on the D-GRU pre-training without sample screening is obtained, and the real-time trajectory slice s is obtained. According to the feature expression method in step (4), s is reduced to a vector (V var,s , A chg,s ), and obtain the predicted value according to the trajectory prediction model based on D-GRU pre-training without sample screening;

[0088] Model C: D-GRU pre-trained trajectory prediction model based on sample screening, as follows:

[0089] Load the D-GRU pre-trained trajectory prediction model based on sample screening obtained in step (65), obtain the real-time trajectory slice s, and reduce s to a vector (V according to the feature expression method in step (4) var,s , A chg,s ), obtain the predicted value according to the loaded D-GRU pre-trained trajectory prediction model based on sample screening;

[0090] Model D: Trajectory prediction model based on particle filtering;

[0091] Model E: UAV trajectory integrated prediction model based on data migration;

[0092] Select models A, B, C, and D as comparison items for model E; use sample set 2 in step (332) as the initial mixed trajectory data slice set in the training phase, and the target drone trajectory slice in the test phase is derived from sample set 1 in step (331). The trajectory prediction accuracy of models A, B, C, and D is obtained respectively, and the accuracy is compared with the drone final trajectory prediction result calculated by model E in step (84) to evaluate the drone trajectory prediction result;

[0093] (922) Sample prediction time measurement;

[0094] Combined with the dimensionality reduction method used in step (4) and the experimental platform used for UAV trajectory prediction, the prediction time consumption of the samples is measured.

[0095] Beneficial effects of the present invention:

[0096] The method of the present invention improves the accuracy of data-driven trajectory prediction of UAVs in tasks with insufficient historical trajectory samples, lays a foundation for flight conflict detection and dangerous behavior identification, and provides effective decision-making support for urban low-altitude airspace management.

[0097] The method of the present invention effectively alleviates the problems of lack of training samples and lengthy training time of data-driven trajectory prediction models by realizing transfer learning of different UAV trajectories and adopting the strategy of model pre-training and real-time prediction, meets the needs of real-time trajectory prediction, and proposes an integrated trajectory prediction method including multiple trajectory prediction models, which increases the generalization ability of the prediction model. BRIEF DESCRIPTION OF THE DRAWINGS

[0098] Figure 1 Flow chart of the method of the present invention. DETAILED DESCRIPTION

[0099] In order to facilitate the understanding of those skilled in the art, the present invention is further described below in conjunction with embodiments and drawings. The contents mentioned in the implementation modes are not intended to limit the present invention.

[0100] Reference Figure 1 As shown, a method for real-time trajectory integration prediction of a UAV based on transfer learning of the present invention comprises the following steps:

[0101] (1) The trajectory data of various types of light rotor UAVs were collected, and the performance parameters of each type of UAV were counted. Table 1 shows some performance parameters of each type of UAV; the performance parameters include net weight, maximum static wind speed, maximum wind speed, maximum climb rate and maximum descent rate; Table 1 is as follows:

[0102] Table 1

[0103]

[0104] (2) Processing the trajectory data of each type of drone collected in the above step (1) into a time series slice data set;

[0105] (2.1) Generate fixed time interval series data;

[0106] Process the trajectory data collected in step (1) into sequence data with a sampling time interval of 5 seconds;

[0107] (2.2) Generate time series slice dataset;

[0108] Use the sliding time window to split the sequence data obtained in step (2.1) into 11 (input sequence length T in The time series slices are 10 and the output length is 1) to generate a time series slice dataset;

[0109] (3) According to the performance parameter constraints of each type of UAV, the time series slice data set of each type of UAV trajectory obtained in step (2) is updated, and the sample set is divided into a predicted target trajectory slice sample set (sample set 1) and a mixed trajectory slice set excluding the predicted target (sample set 2);

[0110] (3.1) Determine speed constraints;

[0111] Solve the vertical component of the estimated velocity of each trajectory point in the time series slice data set obtained in step (2), with the vertical component satisfying the maximum climb / maximum descent rate in the performance parameters of each model of drone as a constraint;

[0112] The estimated velocity V of the i-th trajectory point i It is expressed as:

[0113]

[0114] Where Δt represents the trajectory sampling interval, P i+1 and P i-1 The relative coordinates of the i+1th and i-1th trajectory points relative to the first trajectory point, respectively. The first trajectory point of each trajectory time series slice is set as the coordinate origin on the corresponding slice;

[0115] (3.2) Filtering time series slice datasets;

[0116] According to the flight speed limit (i.e., maximum climb / maximum descent rate constraint) of each type of UAV in step (3.1), the time series slice data set that meets the flight speed limit in step (2) is selected;

[0117] (3.3) Divide the slice sample sets into two different types as follows:

[0118] (3.3.1) Sample set 1: A time series slice dataset using only a single model of the prediction target;

[0119] (3.3.2) Sample set 2: A mixed dataset of time series slices of the trajectories of each type of drone except the predicted target in step (3.2).

[0120] (4) For sample set 1 and sample set 2, trajectory slicing dimensionality reduction based on statistical metrics is performed respectively;

[0121] (4.1) Set the coordinate origin of the trajectory slice;

[0122] The coordinates of the first track point of each track time series slice after division are marked as (0,0,0), that is, set as the coordinate origin on the corresponding slice;

[0123] (4.2) Dimensionality reduction is performed on sample set 1 and sample set 2 respectively based on feature statistics;

[0124] Calculate the speed and heading change of each trajectory point in each trajectory slice relative to the coordinate origin in step (4.1), and then calculate the flight speed variance V of each trajectory slice var and the accumulated heading change A chg :

[0125]

[0126]

[0127] Where, T in is the time step of the input information, V i is the velocity of the ith trajectory point, V s is the distance from the 2nd point to the Tth point in the trajectory slice in -1 point average flight speed, Ai is the heading variable of the i-th trajectory point, and the flight speed variance V var and the accumulated heading change A chg As a potential feature that can comprehensively express time series slices in low-dimensional space, the time series slices in step (3.3) can be expressed in low-dimensional space.

[0128] (5) Screening similar trajectories of training samples for the time series sliced ​​dataset of drone trajectories except for the predicted target after dimensionality reduction;

[0129] (5.1) Quantile division;

[0130] For potential feature V var and A chg , starting from the 0 quantile, m = 3 and n = 3 equally spaced quantiles are additionally set, and the UAV mixed trajectory time series slice dataset except the predicted target after dimensionality reduction is divided into 3 × 3 subsets;

[0131] (5.2) Solve the sample coordinates of the time series slice;

[0132] For any time series slice sample in the mixed sample set in step (5.1), the flight speed variance and cumulative heading change of the sample are used as the potential characteristic flight speed variance V var and the accumulated heading change A chg The position of the quantile point, and the coordinates of the sample are obtained;

[0133] (5.3) Solve the coordinate matrix of the mixed sample set;

[0134] Combined with step (5.2), the coordinates of all time series slice samples in the mixed sample set in step (5.1) are obtained, and the coordinate matrix of the mixed trajectory time series slice data set of the drone except the predicted target is obtained after dimensionality reduction, and then divided into m×n subsets, among which the subsets with large values ​​of m and n will contain small subsets, based on the flight speed variance V var and the accumulated heading change A chg The numerical similarity measures the distance between training samples and performs subset screening on the UAV mixed trajectory time series slice dataset except for the predicted target after dimensionality reduction.

[0135] (6) Using the D-GRU trajectory prediction model to perform model pre-training, and obtaining a D-GRU pre-trained trajectory prediction model based on sample screening;

[0136] (6.1) Setting the D-GRU trajectory prediction model parameters;

[0137] (6.1.1) Set the model input layer parameters;

[0138] Determine the number of training samples in a single batch to be 100, and the input data dimension of each sample to be [10,3];

[0139] (6.1.2) Set the model network layer parameters;

[0140] The model consists of an input layer, three parallel GRU layers, a parallel fully connected layer, and an output layer. The number of neurons in a single layer is 20, and the time step of the input information is 10. The activation function is initially selected as elu.

[0141] (6.1.3) Set the model output layer parameters;

[0142] The model output data dimension is 3; the activation function initially selects elu;

[0143] (6.2) Construct D-GRU trajectory prediction model training samples;

[0144] Take five models (Mavic Air, Mavic 2, Mavic Pro, Spark, Phantom 3) as the prediction targets to obtain slice sets D, and divide D into classification set S and integration set U in a ratio of 7:3; according to steps (4) and (5.3), respectively reduce the dimension of S and U and solve the coordinate matrix; according to step (5.2), divide S into a set S containing 3×3 sample subsets 33 ;

[0145] (6.3) Set the loss function of the D-GRU trajectory prediction model;

[0146] The L2 regularization method is used as the loss function of mean square error (MSE);

[0147] (6.4) Set up the D-GRU trajectory prediction model optimizer;

[0148] Select Adam as the model optimizer with a learning rate of 0.0001;

[0149] (6.5) Train and save the D-GRU trajectory prediction model;

[0150] Take S in step (6.2) 33 The input samples are constructed from each subset of , the D-GRU trajectory prediction model is trained and saved, and the D-GRU pre-trained trajectory prediction model based on sample screening is obtained (model C).

[0151] (7) Integrate the D-GRU pre-trained trajectory prediction model based on sample screening with the trajectory prediction model based on particle filtering to obtain an integrated prediction model;

[0152] (7.1) Solve for each predicted value of the integrated set U;

[0153] For each sample u in the integrated set U in step (6.2), the prediction value p is obtained by combining the D-GRU pre-trained trajectory prediction model (model C) based on sample screening in step (6.5). u,0 , the trajectory prediction model based on particle filtering (model D) obtains the predicted value p u,1 , and get the predicted value p of the integrated set U u,0 and p u,1 ;

[0154] (7.2) Train and save the integrated prediction model;

[0155] The two predicted values ​​of the integrated set U in step (7.1) and its own input data are used as the input of the integrated prediction model, and the true output of the integrated set U is used as the output of the integrated prediction model. The integrated weights of the D-GRU pre-trained trajectory prediction model (model C) based on sample screening and the trajectory prediction model (model D) based on particle filtering in step (7.1) are trained and saved, and the UAV trajectory integrated prediction model (model E) based on data migration is obtained.

[0156] (8) Real-time prediction of target trajectory slices based on the integrated prediction model;

[0157] (8.1) Load the pre-trained model;

[0158] Load the D-GRU pre-trained trajectory prediction model (model C) based on sample screening in step (6.5) and the drone trajectory integration prediction model (model E) based on data migration in step (7.2); obtain the real-time trajectory slice s;

[0159] (8.2) Real-time trajectory slicing dimensionality reduction;

[0160] According to the feature expression method in step (4), s is reduced to a vector (V var,s , A chg,s );

[0161] (8.3) Solve the prediction values ​​of the D-GRU pre-trained trajectory prediction model based on sample screening (model C) and the trajectory prediction model based on particle filtering (model D) respectively;

[0162] The predicted value p is obtained according to the D-GRU pre-trained trajectory prediction model (model C) based on sample screening loaded in step (8.1) s,0 , the trajectory prediction model based on particle filtering (model D) obtains the predicted value p s,1 ;

[0163] (8.4) Solve the prediction value of the integrated prediction model;

[0164] Based on the data migration-based UAV trajectory integrated prediction model (model E) loaded in step (8.1), the final prediction value p of the trajectory slice s is obtained s .

[0165] (9) Evaluate the UAV trajectory prediction results;

[0166] (9.1) Calculate trajectory prediction accuracy;

[0167] According to the final predicted values ​​of each type of drone obtained in step (8.4), calculate the average Euclidean error E between the predicted coordinates and the true coordinates in three-dimensional space e , which is used as an evaluation index to evaluate the accuracy of UAV trajectory prediction;

[0168] (9.2) Prediction accuracy comparison and time measurement;

[0169] (9.2.1) Comparison of the accuracy of different prediction models;

[0170] Model A: D-GRU trajectory prediction model based on predicted target trajectory data, as follows:

[0171] For the sample set 1 in the step (3.3.1) after dimensionality reduction, i.e., the time series slice data set for predicting the target trajectory, the trajectory prediction model based on D-GRU is used to obtain the predicted value;

[0172] Model B: Trajectory prediction model based on D-GRU pre-training without sample screening, as follows:

[0173] When constructing the training samples of the D-GRU trajectory prediction model in step (6.2), no sample screening is performed, that is, m=n=1 is set when the classification set S is divided. According to step (6), the trajectory prediction model based on the D-GRU pre-training without sample screening is obtained, and the real-time trajectory slice s is obtained. According to the feature expression method in step (4), s is reduced to a vector (V var,s , A chg,s ), and obtain the predicted value according to the trajectory prediction model based on D-GRU pre-training without sample screening;

[0174] Model C: D-GRU pre-trained trajectory prediction model based on sample screening, as follows:

[0175] Load the D-GRU pre-trained trajectory prediction model based on sample screening obtained in step (6.5), obtain the real-time trajectory slice s, and reduce s to a vector (V according to the feature expression method in step (4) var,s , A chg,s ), obtain the predicted value according to the loaded D-GRU pre-trained trajectory prediction model based on sample screening;

[0176] Model D: The trajectory prediction model based on particle filtering obtains the predicted value;

[0177] Model E: UAV trajectory integrated prediction model based on data migration;

[0178] Select models A, B, C, and D as comparison items for model E; use sample set 2 in step (3.3.2) as the initial mixed trajectory data slice set in the training phase, and the target drone trajectory slice in the test phase is derived from sample set 1 in step (3.3.1). The trajectory prediction accuracy of models A, B, C, and D for 5s, 10s, 15s, and 20s is obtained respectively, and the accuracy is compared with the final trajectory prediction result of the drone calculated by model E in step (8.4) to evaluate the drone trajectory prediction result. As shown in Table 2, the average Euclidean error of trajectory prediction under different prediction durations, the method proposed in the present invention has excellent performance in the real-time trajectory prediction method for drones that lack sufficient real historical trajectories compared to the trajectory prediction model that does not perform trajectory screening; Table 2 is as follows:

[0179] Table 2

[0180]

[0181] (9.2.2) Sample prediction time measurement;

[0182] Based on the pre-training method of the model, combined with the dimensionality reduction method used in step (4) and the Ryzen R5 3600CPU and NVIDIA RTX2060 GPU experimental platform, the prediction time of the samples was measured, and the average prediction time of multiple samples was 0.1476 seconds, which can meet the real-time requirements.

[0183] The present invention has many specific application paths. The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements can be made without departing from the principle of the present invention. These improvements should also be regarded as the protection scope of the present invention.

Claims

1. A real-time trajectory integrated prediction method for unmanned aerial vehicles based on transfer learning, characterized in that: Here are the steps: (1) Collect trajectory data of various types of light rotary-wing UAVs and calculate the performance parameters of each type of UAV; (2) Processing the trajectory data of each type of drone collected in the above step (1) into a time series slice data set; (3) According to the performance parameter constraints of each type of UAV, the time series slice data set of each type of UAV trajectory obtained in step (2) is updated, and the sample set is divided into the time series slice data set of the UAV trajectory of the predicted target and the time series slice data set of the UAV trajectory other than the predicted target; (4) For the UAV trajectory time series slice dataset of the predicted target and the UAV trajectory time series slice dataset other than the predicted target, trajectory slice dimension reduction based on statistical measurement is performed respectively; (5) Screening similar trajectories of training samples for the time series sliced ​​dataset of drone trajectories except for the predicted target after dimensionality reduction; (6) Using the D-GRU trajectory prediction model to perform model pre-training, and obtaining a D-GRU pre-trained trajectory prediction model based on sample screening; (7) Integrate the D-GRU pre-trained trajectory prediction model based on sample screening with the trajectory prediction model based on particle filtering to obtain an integrated prediction model; (8) Real-time prediction of target trajectory slices based on the integrated prediction model; (9) Evaluate the UAV trajectory prediction results; The specific process of step (3) is as follows: (31) Determine speed constraints; Solve the vertical component of the estimated velocity of each trajectory point in the time series slice data set obtained in step (2), with the vertical component satisfying the maximum climb / maximum descent rate in the performance parameters of each model of drone as a constraint; The estimated velocity V of the i-th trajectory point i It is expressed as: Where Δt represents the trajectory sampling interval, P i+1 and P i-1 The relative coordinates of the i+1th and i-1th trajectory points relative to the first trajectory point, respectively. The first trajectory point of each trajectory time series slice is set as the coordinate origin on the corresponding slice; (32) Filtering time series slice datasets; According to the flight speed limit of each type of UAV in step (31), the time series slice data set that meets the flight speed limit in step (2) is selected; (33) Two different types of slice sample sets are divided as follows: (331) Sample set 1: The time series slice dataset of the drone trajectory of the predicted target in step (32); (332) Sample set 2: Mix the drone trajectory time series slice datasets except the predicted target model in step (32) to obtain sample set 2; The specific process of step (4) is as follows: (41) Setting the coordinate origin of the trajectory slice; The coordinates of the first track point of each track time series slice after division are marked as (0,0,0), that is, set as the coordinate origin on the corresponding slice; (42) Based on feature statistics, sample set 1 and sample set 2 are reduced in dimension respectively; The speed and heading change of each trajectory point in each trajectory slice relative to the coordinate origin in step (41) are calculated respectively, and then the flight speed variance V of each trajectory slice is calculated. var and the accumulated heading change A chg , taking it as a potential feature that can comprehensively express the time series slices in low-dimensional space, so that the time series slices in step (33) are expressed in low-dimensional space; The specific process of step (5) is as follows: (51) Quantile division; For the potential feature flight speed variance V var and the accumulated heading change A chg , starting from the 0 quantile, m and n equally spaced quantiles are set separately, and the UAV mixed trajectory time series slice dataset except the predicted target after dimensionality reduction is divided into m×n subsets; (52) Solve the sample coordinates of the time series slice; For any time series slice sample in the mixed sample set in step (51), based on the flight speed variance and cumulative heading change of the sample, the potential characteristic flight speed variance V var and the accumulated heading change A chg The position of the quantile point, and the coordinates of the sample are obtained; (53) Solve the coordinate matrix of the mixed sample set; Combined with step (52), the coordinates of all time series slice samples in the mixed sample set in step (51) are obtained, and the coordinate matrix of the mixed trajectory time series slice data set of the drone except the predicted target after dimensionality reduction is obtained, and then divided into m×n subsets, among which the subset with large values ​​of m and n will contain the small subset, based on the flight speed variance V var and the accumulated heading change A chg The similarity of values,measures the distance between training samples, and performs subset screening on the,drone mixed trajectory time series slice dataset except the predicted target after,dimension reduction; The specific process of step (6) is as follows: (61) Setting the D-GRU trajectory prediction model parameters; (611) Setting model input layer parameters; Determine the number of training samples in a single batch and the input data dimension of each sample; initialize the input layer weights and biases; (612) Setting model network layer parameters; Determine the number of hidden layers, the number of neurons in a single layer, and the time step of input information; initialize the network layer weights and biases; initially select elu as the activation function; (613) Setting model output layer parameters; Determine the output data dimension; initialize the output layer weights and biases; initially select elu as the activation function; (62) Construct D-GRU trajectory prediction model training samples; The sample set 2 in step (332) is divided into a classification set S and an integration set U in a ratio of 7:3; according to steps (4) and (53), S and U are respectively reduced in dimension and the coordinate matrix is ​​solved, and then S is divided into a set S containing m×n sample subsets mn ; (63) Setting the loss function of the D-GRU trajectory prediction model; The L2 regularization method is used as the loss function of the mean square error; (64) Setting the D-GRU trajectory prediction model optimizer; Select Adam as the model optimizer with a learning rate of 0.0001; (65) Train and save the D-GRU trajectory prediction model; Take the set S in step (62) mn Construct input samples from each subset, train the D-GRU trajectory prediction model and save it, and obtain the D-GRU pre-trained trajectory prediction model based on sample screening; The specific process of step (7) is as follows: (71) Solve for each predicted value of the integrated set U; For each sample u in the integrated set U in step (62), the prediction value p is obtained by combining the D-GRU pre-trained trajectory prediction model based on sample screening in step (65). u,0 , the trajectory prediction model based on particle filtering obtains the predicted value p u,1 , and get the predicted value p of the integrated set U u,0 and p u,1 ; (72) Integrated prediction model training and storage; The two predicted values ​​of the integrated set U in step (71) and its own input data are used as the input of the integrated prediction model, and the real output of the integrated set U is used as the output of the integrated prediction model. The integrated weights of the trajectory prediction model pre-trained by D-GRU based on sample screening and the trajectory prediction model based on particle filtering in step (71) are trained and saved, and the integrated prediction model is obtained to obtain the UAV trajectory integrated prediction model based on data migration; The specific process of step (8) is as follows: (81) Loading the pre-trained model; Load the D-GRU pre-trained trajectory prediction model based on sample screening in step (65) and the drone trajectory integration prediction model based on data migration in step (72); obtain the real-time trajectory slice s; (82) Real-time trajectory slicing dimensionality reduction; According to the feature expression method in step (4), s is reduced to a vector (V var,s , A chg,s ); (83) respectively solve the prediction values ​​of the trajectory prediction model based on sample screening D-GRU pre-training and the trajectory prediction model based on particle filtering; The predicted value p is obtained according to the D-GRU pre-trained trajectory prediction model loaded based on sample screening in step (81) s,0 , the trajectory prediction model based on particle filtering obtains the predicted value p s,1 ; (84) solving the prediction value of the integrated prediction model; Based on the data migration-based UAV trajectory integrated prediction model loaded in step (81), the final prediction value p of the trajectory slice s is obtained s ; The specific process of step (9) is as follows: (91) Calculate trajectory prediction accuracy; According to the final predicted values ​​of each type of drone obtained in step (84), the average Euclidean error E between the predicted coordinates and the true coordinates in three-dimensional space is calculated. e , which is used as an evaluation index to evaluate the accuracy of UAV trajectory prediction; (92) Comparison of prediction accuracy and time consumption measurement; (921) Comparison of the accuracy of different prediction models; Model A: D-GRU trajectory prediction model based on predicted target trajectory data, as follows: For the sample set 1 in the step (331) after dimensionality reduction, i.e., the time series slice data set for predicting the target trajectory, the prediction result is obtained using the trajectory prediction model based on D-GRU; Model B: Trajectory prediction model based on D-GRU pre-training without sample screening, as follows: When constructing the training samples of the D-GRU trajectory prediction model in step (62), no sample screening is performed, that is, m=n=1 is set when the classification set S is divided. According to step (6), the trajectory prediction model based on the D-GRU pre-training without sample screening is obtained, and the real-time trajectory slice s is obtained. According to the feature expression method in step (4), s is reduced to a vector (V var,s , A chg,s ), and obtain the predicted value according to the trajectory prediction model based on D-GRU pre-training without sample screening; Model C: D-GRU pre-trained trajectory prediction model based on sample screening, as follows: Load the D-GRU pre-trained trajectory prediction model based on sample screening obtained in step (65), obtain the real-time trajectory slice s, and reduce s to a vector (V according to the feature expression method in step (4) var,s , A chg,s ), obtain the predicted value according to the loaded D-GRU pre-trained trajectory prediction model based on sample screening; Model D: Trajectory prediction model based on particle filtering; Model E: UAV trajectory integrated prediction model based on data migration; Select models A, B, C, and D as comparison items for model E; use sample set 2 in step (332) as the initial mixed trajectory data slice set in the training phase, and the target drone trajectory slice in the test phase is derived from sample set 1 in step (331). The trajectory prediction accuracy of models A, B, C, and D is obtained respectively, and the accuracy is compared with the drone final trajectory prediction result calculated by model E in step (84) to evaluate the drone trajectory prediction result; (922) Sample prediction time measurement; Combined with the dimensionality reduction method used in step (4) and the experimental platform used for UAV trajectory prediction, the prediction time consumption of the samples is measured.

2. The method for real-time trajectory integration prediction of unmanned aerial vehicles based on transfer learning according to claim 1 is characterized in that: In step (1), each track data includes multiple track points, and the information recorded at each track point includes: a universal coordinated timestamp, a drone ID, longitude, latitude, flight altitude, and flight ground speed.

3. The method for real-time trajectory integration prediction of unmanned aerial vehicles based on transfer learning according to claim 1 is characterized in that: The specific process of step (2) is as follows: (21) Generate equally spaced time series data; The trajectory data of each type of UAV collected in step (1) are processed into sequence data with equal time intervals through linear interpolation; (22) Generate a time series slice dataset; A sliding time window is used to divide the sequence data of each equal time interval obtained in step (21) into time series slices of fixed length, thereby generating a time series slice dataset that contains both input and output data in the trajectory prediction model.

Images