A Trajectory Prediction Method Based on Generative Adversarial Network and Long Short-Term Memory Network

Abstract

The present invention discloses a multi-intention trajectory prediction method based on a generative adversarial network and a long short-term memory network, belonging to the technical field of artificial intelligence time series prediction, including: for the time series prediction problem, a multi-intention trajectory prediction method based on a generative adversarial network and a long short-term memory network is proposed. Aiming at the problems of missing trajectory data, long-term dependence, gradient disappearance, limited training data, weak real-time performance and generalization ability in the trajectory prediction task, small sample data augmentation and small sample intention recognition are proposed, as well as a trajectory prediction method for trajectory completion using the spatio-temporal correlation and historical information between data, while ensuring the rapidity, robustness and accuracy of the trajectory prediction algorithm; a GAN network capable of expanding the original data and an LSTM network for trajectory completion are constructed.

Description

Technical Field

[0001] The present invention relates to the field of aerospace, and particularly to a flight track prediction method based on a generative adversarial network and a long short-term memory network. Background Art

[0002] The progress of aviation technology has made the aircraft trajectory completion algorithm a research hotspot in the aviation field. This algorithm aims to solve the problem of missing trajectory data caused by various reasons during flight, which is crucial for flight safety and data accuracy. Currently, interpolation methods, such as linear interpolation, polynomial interpolation, and spline interpolation, are widely used as basic technologies. They infer and complete trajectory data by establishing a mathematical model between known points. However, due to the complexity and variability of the aircraft trajectory, these methods face challenges in dealing with non-linear and complex systems.

[0003] To overcome the limitations of existing methods, researchers have proposed a variety of new algorithms. For the problem of missing ADS-B data, a multi-constraint discriminant model is established to screen missing points, and an improved proximity algorithm and multiple imputation method are used for trajectory repair. In addition, based on the dynamic change of the heading angle, researchers have proposed an interpolation fitting completion algorithm for missing trajectory points. Sun Lishuang et al. proposed a neighborhood equal ratio filling algorithm, which uses the continuity characteristics of neighborhood trajectories to determine the position of filling points. Yu et al. studied a discrete Kalman filter trajectory prediction algorithm based on neural accelerometer estimators, which is applicable to the trajectory prediction and completion of military aircraft. In addition, deep learning methods, such as Bi-LSTM and Transformer, have also been applied to trajectory completion, using their non-linear processing ability and self-attention mechanism to improve prediction accuracy.

[0004] Although the existing research has made progress, the field of trajectory completion still faces some challenges. The method based on the physical model requires a large amount of prior knowledge and experience, and it is difficult to adapt to the complex and changeable actual situation. Although deep learning algorithms provide new solutions, the problem of small sample learning is still a difficult problem because it is very difficult to obtain sufficient samples in practice. Currently, the research on small sample learning focuses on three aspects: data augmentation, hypothesis space constraint, and search strategy adjustment, aiming to expand the small sample data set through prior knowledge and improve the generalization ability and credibility of the model. However, how to systematically design small sample learning strategies and how to deal with the problem of inconsistency of model structure and parameters caused by too few samples are still difficult points that need to be further explored in the research. Summary of the Invention

[0005] The purpose of the present invention is to overcome the defects existing in the prior art and solve the problem of predicting the missing aircraft trajectory under the condition of small samples.

[0006] To achieve the above object, the present invention proposes a trajectory prediction method based on a generative adversarial network and a long short-term memory network, including the following steps:

[0007] S1. Read the collected data and perform data cleaning.

[0008] S2. Perform polynomial fitting interpolation for the missing trajectory data. Since there is a certain linear correlation between the trajectory vector and the time series, the known data can be used to construct a fitting polynomial to fill in the missing trajectory data. The fitting polynomial is:

[0009]

[0010] where Y m is the predicted value of the dependent variable, α0 is the intercept term, α i is the regression coefficient, X i is the value of the i-th independent variable in X, and N is the total number of data.

[0011] S3. Normalize the three-dimensional trajectory feature data. The normalization formula for mapping the original data to the interval [0, 1] is:

[0012]

[0013] where the subscript x is the x-th dimensional data in the trajectory vector, the subscript i is the i-th data, f′ is the value after the mapping transformation, f is the value before the transformation, f′ xi represents the value of the i-th original data f xi in the x-th dimension after normalization. F is the set of all vectors before the transformation, that is, F = {f1, f2,..., f i ,..., f N}, F x is the x-th dimensional data in the set of all vectors before the transformation, and N is the total number of data.

[0014] S4. Process the small sample data set. The specific process is as follows: First, divide the small sample data set according to the intention, and extract the trajectories with the same intention and keep their time order into a new data set, so that the divided data set only contains data of a single type of intention.

[0015] S5. Construct a TimeGAN network. The TimeGAN network includes five parts, namely a generation network, an embedding network, a recovery network, a discriminant network, and a supervision network. The structure of each network is composed of a 24-layer recurrent neural network plus a fully connected layer.

[0016] S6. Construct a loss function and perform embedding training and supervised training in sequence. Through the training of these four networks, the data can maintain a distribution similar to the original after passing through the embedding network and the recovery network. First, give the output representations of the input X and the noise Z after passing through different networks, where the noise Z is a random noise uniformly distributed from 0 to 1 with the same shape as the input data X. The output of the input X passing through the embedding network is denoted as H, the output of H passing through the recovery network is denoted as denoted as, the output of H passing through the supervision network is denoted as denoted as, the output of H passing through the discrimination network is denoted as Y real denoted as; the output of the noise Z passing through the generation network is denoted as denoted as, the output passing through the supervision network is denoted as denoted as, the output passing through the recovery network is denoted as denoted as, the output passing through the discrimination network is denoted as Y fake denoted as, the output passing through the discrimination network is denoted as Y efake denoted as.

[0017] And store all the parameters of the five networks into a list respectively. The parameter list of the embedding network is denoted as e v denoted as, the parameter list of the recovery network is denoted as e r denoted as, the parameter list of the generation network is denoted as e g denoted as, the parameter list of the supervision network is denoted as e s denoted as, the parameter list of the discrimination network is denoted as e d denoted as.

[0018] The specific process is as follows: Calculate the mean square loss during the embedding training process:

[0019]

[0020] where N is the total number of data, the subscript i is the i-th component in the data, X i is the i-th component in the input data, is the i-th component in. Use the mean square error to evaluate the difference between the two, and use the adam optimizer to update the parameters of e loss

[0021] where θ1 is the parameter to be updated, including e v and e r , α represents the learning rate, which is a hyperparameter used to control the step size of each update. is an unbiased estimate of the gradient, used to guide the direction of parameter update.​ is the variance of the unbiased estimate of the gradient. ∈ is a very small positive number used for numerical stability to prevent the denominator from being zero.

[0022] Calculate the mean square loss during the supervised training process:

[0023]

[0024] where N is the total number of data, the subscript i is the i-th component in the data, and X i is the i-th component in the input data, is the i-th component in. Use the mean square error between the two as the loss of the supervised network, and use the adam optimizer to update the parameters of g loss Update the parameters

[0025] where θ2 are the parameters to be updated, including e g and e s , and α represents the learning rate, which is a hyperparameter used to control the step size of each update. is the unbiased estimate of the gradient, which is used to guide the direction of parameter update. is the variance of the unbiased estimate of the gradient. ∈ is a very small positive number used for numerical stability to prevent the denominator from being zero.

[0026] S7. Construct the joint loss function and perform joint training. First, calculate the cross-entropy loss between Y fake and the all-1 vector of the same shape as Y fake , Secondly, calculate the cross-entropy loss between Y efake and the all-1 vector of the same shape as Y efake , where is used to map any real number to the interval (0, 1), N is the total number of samples, Y fake,i is the predicted value of the i-th sample in Y fake (i.e., the output of the model), and Y efake,i is the predicted value of the i-th sample in Y efake . Calculate G Gloss = g loss +g lossfr +g lossfre , and use the adam optimizer to update the parameters of G Gloss Update the parameters

[0027] where θ3 are the parameters to be updated, including e g and e s , and α represents the learning rate, which is a hyperparameter used to control the step size of each update. Is an unbiased estimate of the gradient and is used to guide the direction of parameter update. Is the variance of the unbiased estimate of the gradient. ∈ is a very small positive number used for numerical stability to prevent the denominator from being zero.

[0028] Calculate Y real And the cross-entropy loss with the all-ones vector of the same shape as Y real Next, calculate Y fake And the cross-entropy loss with the all-zeros vector of the same shape as Y fake Next, calculate Y efake And the cross-entropy loss with the all-zeros vector of the same shape as Y efake Where Is used to map any real number to the interval (0, 1). N is the total number of samples. Y fake,i Is fake The predicted value of the i-th sample in Y (i.e., the output of the model). Y efake,i Is efake The predicted value of the i-th sample in Y. Calculate G Dloss = g lossrr + g lossff + g lossffe , and use the adam optimizer to update the parameters of G Dloss Update the parameters

[0029] Where θ4 is the parameter to be updated, including e d , α represents the learning rate, which is a hyperparameter used to control the step size of each update. Is an unbiased estimate of the gradient and is used to guide the direction of parameter update. Is the variance of the unbiased estimate of the gradient. ∈ is a very small positive number used for numerical stability to prevent the denominator from being zero.

[0030] In each training batch, first update θ3 twice and then update θ4 once.

[0031] S8. Construct the hidden layer LSTM network to predict the trajectory. The specific process is as follows: Construct the forget gate, update gate, and output gate through the following formula:

[0032] Γ f = σ(W f [h t-1 , x t + b f )

[0033] Γ u = σ(W u [h t-1 , x t+b u )

[0034] Γ a =σ(W a [h t-1 ,x t +b a )

[0035] where Γ f , Γ u and Γ a represent the forget gate, update gate, and output gate respectively, σ is the Sigmoid function, W f , W u and W a are the weight coefficient matrices for the corresponding parts, h is the output feature, x is the input feature, the subscript t represents the current time step, t - 1 represents the previous time step, and b f , b u and b a are the offset vectors for the corresponding parts. The candidate neuron is calculated by the following formula:

[0036]

[0037] where W c is the weight coefficient matrix for this part, and b c is the offset vector for this part. The updated neuron is calculated by the following formula:

[0038]

[0039] The output feature at the current time step is calculated by the following formula:

[0040] h t =Γ a *tanh C t

[0041] S9. Construct the output fully connected layer to obtain the three - dimensional coordinate prediction trajectory. The specific process is as follows: The fully connected layer is constructed by the following formula:

[0042] y f =W f ·h t +b f

[0043] where W f is the weight matrix of the fully connected layer, b f is the bias of the fully connected layer, and y f is the output scalar of the fully connected layer, containing the prediction trajectory information of one time step. During prediction, one trajectory point of one dimension at one time step is predicted each time, and finally, the three - dimensional coordinate prediction trajectory corresponding to one time step is output.

[0044] Compared with the prior art, the advantages and beneficial effects of the present invention are as follows: Aiming at the problems of missing trajectory data, long-term dependence, vanishing gradient, limited training data, weak real-time performance and generalization ability in the trajectory prediction task, a small-sample data augmentation and small-sample intention recognition, as well as a trajectory prediction method for trajectory completion using the spatio-temporal correlation and historical information between data are proposed, while ensuring the rapidity, robustness and accuracy of the trajectory prediction algorithm; A GAN network capable of expanding the original data and an LSTM network for trajectory completion are constructed. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 It is a flow chart of a trajectory prediction method based on a generative adversarial network and a long short-term memory network of the present invention.

[0046] Figure 2 It is an operation structure diagram of the LSTM network of the present invention.

[0047] Figure 3 It is a model structure diagram of the LSTM network of the present invention.

[0048] Figure 4 It is a prediction result diagram of the original data of an example of the present invention.

[0049] Figure 5 It is a prediction result diagram of the augmented data of an example of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0050] To further illustrate the features of the present invention, please refer to the following detailed description of the present invention and the attached Figures 1-5 . The attached drawings are for reference and illustration only, and are not intended to limit the protection scope of the present invention.

[0051] The following will describe in detail the specific embodiments of the present invention with reference to the attached drawings:

[0052] As Figure 1 shown, this embodiment discloses a trajectory prediction method based on a generative adversarial network and a long short-term memory network of the present invention, including the following steps:

[0053] S1. Read the collected data and perform data cleaning. Data collection part: FlightADSB is a platform that provides real-time tracking services for global flight trajectories. Users can obtain information services such as real-time tracking maps of aircraft, playback of aircraft flight trajectories, download of flight trajectory data, flight status, application for using ADS-B devices, and display of aircraft pictures through FlightADSB. Therefore, in the example part of the present invention, three flight trajectories in FlightADSB corresponding to three intentions are selected. The intention label mapping of flight number HO1702 (Guiyang Longdongbao - Nanjing Lukou) is 1, the intention label mapping of CZ6796 (Guiyang Longdongbao - Sanya Phoenix) is 2, and the intention label mapping of ZH8210 (Guiyang Longdongbao - Shenzhen Bao'an) is 3. 10 complete trajectories and 10 incomplete trajectories are taken from each as the training data set to reflect the effect of this method in solving the small sample problem under the condition of insufficient data and data loss. In addition, in order to reflect the limitations of real conditions and current sensor technologies, the data that can be obtained in real time is retained, including flight timestamps, geographical coordinates, altitude and other information. Data cleaning part: It includes operations such as deleting duplicate values, performing consistency processing, and detecting and handling outliers.

[0054] S2. Perform polynomial fitting interpolation for the missing data in the trajectory. In the real scenario, data may be lost during the processes of collection, transmission, and analysis, and the integrity of the data set is crucial for accurately identifying the true intentions of flight targets. Therefore, it is very necessary to effectively fill the missing parts in the trajectory. Since there is a certain linear correlation between the trajectory vectors and the time series, the known data can be used to construct a fitting polynomial to fill the missing trajectory data. The fitting polynomial is:

[0055]

[0056] where Y m is the data instance X in the data set that contains one or more missing attribute values mis ,

[0057] X i (i = 1, …, N) are the instances X in the data set without any missing attribute values obs .

[0058] S3. Normalize the three-dimensional trajectory feature data. Data normalization can eliminate the influence of data dimensions and improve the network convergence efficiency. The normalization formula for mapping the original data to the interval [0, 1] is:

[0059]

[0060] Among them, the subscript x represents the x - dimensional data in the trajectory vector, the subscript i represents the i - th data, f′ is the value after the mapping transformation, f is the value before the transformation, F is the set of all vectors before the transformation, that is, F = {f1, f2, …, f i , …, f N}, and N is the total number of data.

[0061] S4. Process the small - sample data set. The specific process is as follows: First, segment the small - sample data set according to the intention, extract the tracks with the same intention in their time order into a new data set, so that the segmented data set only contains data of a single - class intention.

[0062] S5. Construct the TimeGAN network. The TimeGAN network consists of five parts, namely the generation network, the embedding network, the recovery network, the discriminator network, and the supervision network. The structure of each network is composed of a 24 - layer recurrent neural network plus a fully - connected layer.

[0063] S6. Construct the loss function and conduct embedding training and supervision training in sequence. Through the training of these four networks, the data can maintain a distribution similar to the original after passing through the embedding network and the recovery network. First, give the output representations of the input X and the noise Z after passing through different networks, where the noise Z is a random noise with a uniform distribution from 0 to 1 and having the same shape as the input data X. The output of the input X passing through the embedding network is denoted as H, the output of H passing through the recovery network is denoted as , the output of H passing through the supervision network is denoted as , the output of H passing through the discriminator network is denoted as Y real ; the output of the noise Z passing through the generation network is denoted as , the output of passing through the supervision network is denoted as , the output of passing through the recovery network is denoted as , the output of passing through the discriminator network is denoted as Y fake , the output of passing through the discriminator network is denoted as Y efake .

[0064] And store all the parameters of the five networks into a list respectively. The parameter list of the embedding network is denoted as e v , the parameter list of the recovery network is denoted as e r , the parameter list of the generation network is denoted as e g , the parameter list of the supervision network is denoted as e s , and the parameter list of the discriminator network is denoted as e d .

[0065] The specific process is as follows: During the embedding training process, calculate the mean square loss:

[0066]

[0067] where N is the total number of data, the subscript i is the i-th component in the data, X i is the i-th component in the input data, is the i-th component in. Use the mean square error to evaluate the difference between the two, and adopt the adam optimizer to update the parameter loss

[0068] where θ1 is the parameter to be updated, including e v and e r , α represents the learning rate, which is a hyperparameter used to control the step size of each update. is an unbiased estimate of the gradient, which is used to guide the direction of parameter update. is the variance of the unbiased estimate of the gradient. ∈ is a very small positive number used for numerical stability to prevent the denominator from being zero.

[0069] During the supervised training process, calculate the mean square loss:

[0070]

[0071] where N is the total number of data, the subscript i is the i-th component in the data, X i is the i-th component in the input data, is the i-th component in. Use the mean square error between the two as the loss of the supervised network, and adopt the adam optimizer to update the parameter loss

[0072] where θ2 is the parameter to be updated, including e g and e s , α represents the learning rate, which is a hyperparameter used to control the step size of each update. is an unbiased estimate of the gradient, which is used to guide the direction of parameter update. is the variance of the unbiased estimate of the gradient. ∈ is a very small positive number used for numerical stability to prevent the denominator from being zero.

[0073] S7. Construct the joint loss function and perform joint training of the network. First, calculate the cross-entropy loss between Y fake and the all-1 vector with the same shape as Y fake , Secondly, calculate the cross-entropy loss between Y efake and the all-1 vector with the same shape as Y efake ​​Cross-entropy loss of all-1 vectors of the same shape where is used to map any real number to the interval (0, 1), N is the total number of samples, and Y fake,i is the predicted value of the i-th sample in Y fake (i.e., the output of the model), and Y efake,i is the predicted value of the i-th sample in Y efake Calculate G Gloss = g loss + g lossfr + g lossfre and use the adam optimizer to update the parameters of G Gloss Update parameters

[0074] where θ3 is the parameter to be updated, including e g and e s , and α represents the learning rate, which is a hyperparameter used to control the step size of each update. is an unbiased estimate of the gradient and is used to guide the direction of parameter update. is the variance of the unbiased estimate of the gradient. ∈ is a very small positive number used for numerical stability to prevent the denominator from being zero.

[0075] Calculate the cross-entropy loss of Y real and all-1 vectors of the same shape as Y real Cross-entropy loss of all-1 vectors of the same shape

[0076] Secondly, calculate the cross-entropy loss of Y fake and all-0 vectors of the same shape as Y fake Cross-entropy loss of all-0 vectors of the same shape Secondly, calculate the cross-entropy loss of Y efake and all-0 vectors of the same shape as Y efake Cross-entropy loss of all-0 vectors of the same shape where is used to map any real number to the interval (0, 1), N is the total number of samples, and Y fake,i is the predicted value of the i-th sample in Y fake (i.e., the output of the model), and Y efake,i is the predicted value of the i-th sample in Y efake Calculate G Dloss = g lossrr + g lossff + g lossffe and use the adam optimizer to update the parameters of G Dloss Update parameters

[0077] where θ4 is the parameter to be updated, including e d, α represents the learning rate, which is a hyperparameter used to control the step size of each update. is an unbiased estimate of the gradient and is used to guide the direction of parameter updates. is the variance of the unbiased estimate of the gradient. ∈ is a very small positive number used for numerical stability to prevent the denominator from being zero.

[0078] In each training batch, θ3 is updated twice first, and then θ4 is updated once.

[0079] S8. Construct the hidden layer LSTM network to predict the trajectory. The specific process is as follows: Construct the forget gate, update gate, and output gate through the following formula:

[0080] Γ f = σ(W f [h t-1 , x t + b f )

[0081] Γ u = σ(W u [h t-1 , x t + b u )

[0082] Γ a = σ(W a [h t-1 , x t + b a )

[0083] where Γ f , Γ u and Γ a represent the forget gate, update gate, and output gate respectively, σ is the Sigmoid function, W f , W u and W a are the weight coefficient matrices for the corresponding parts, h is the output feature, x is the input feature, the subscript t represents the current moment, t - 1 represents the previous moment, b f , b u and b a are the offset vectors for the corresponding parts. Calculate the candidate neurons through the following formula:

[0084]

[0085] where W c is the weight coefficient matrix for this part, b c is the offset vector for this part. Calculate the updated neurons through the following formula:

[0086]

[0087] Calculate the output feature at the current moment by the following formula:

[0088] h t = Γ a *tanh C t

[0089] The operation structure diagram of the above LSTM network is as Figure 2 shown. The LSTM model is divided into three parts: the historical trajectory input layer, the hidden layer, and the output layer. The input node of the network is input=(100,3), and the output node is output = 3, where 100 represents the historical trajectory step length; 3 represents the feature dimension. The structure diagram of the LSTM network model is as Figure 3 shown.

[0090] S9. Construct the output fully connected layer to obtain the three-dimensional coordinate prediction trajectory. The specific process is as follows: Construct the fully connected layer by the following formula:

[0091] y f = W f ·h t + b f

[0092] where W f is the weight matrix of the fully connected layer, b f is the bias of the fully connected layer, and y f is the output scalar of the fully connected layer, containing the prediction trajectory information of one time step. When predicting, the trajectory point of one time step of one dimension is predicted each time, and finally the three-dimensional coordinate prediction trajectory corresponding to one time step is output.

[0093] Those of ordinary skill in the art can understand that all or part of the steps of implementing the above method embodiments can be completed by hardware related to program instructions. The foregoing program can be stored in a computer-readable storage medium. When the program is executed, it executes the steps including the above method embodiments; and the foregoing storage medium includes: various media such as ROM, RAM, magnetic disk, or optical disc that can store program codes.

[0094] The following is a simulation example. The present invention will give examples to illustrate the effects of the proposed method. This case shows the advantages of the proposed method.

[0095] First, map the three flight tracks in FlightADSB to three intention labels respectively. HO17002 is mapped to intention Figure 1 , CZ6796 is mapped to intention Figure 2 , and ZH8210 is mapped to intention Figure 3 . Secondly, construct TimeGAN to expand the generated data.

[0096] Perform principal component analysis on the generated results in the example using the PCA package in Scikit-Learn. The pca.components_ attribute of the PCA model gives the eigenvectors corresponding to each principal component. For dimensionality reduction, usually only a part of the principal components are selected to be retained, which means that not all variances are considered. The principal component ratios and their corresponding original features for various intents are shown in the following table. The numbers outside the parentheses in the table represent the number of the feature corresponding to the data of the principal component, and the numbers in the parentheses represent the contribution size and direction of each original feature in this principal component.

[0097] Table 1 Principal Component Ratios and Features of TGAN for Different Intents

[0098]

[0099] It is observed in Table 1 that the principal components of the generated data show significant similarities in features with the principal components of the original data. The principal components of the generated data are similar to those of the original data in terms of direction and proportion, which indicates that the generated data retains features similar to the original data in the direction of the principal components. This property shows that the generated data by the generation algorithm is consistent with the distribution of the original data in key aspects.

[0100] Combine the augmented data with the original data and add intent labels according to different intents. Feed it into the LSTM network for training the trajectory prediction model. The predicted trajectory of the original data set is as Figure 4 shown, and the RMSE value between the original trajectory and the predicted trajectory is 79.687. The predicted trajectory of the augmented data set is as Figure 5 shown, and the RMSE value between the original trajectory and the predicted trajectory is 64.189. Compared with the original data set, the method adopted in the present invention improves the accuracy of trajectory prediction on three-dimensional data.

Claims

1. A track prediction method based on generative adversarial network and long short-term memory network, characterized in that: The steps include: S1, read the collected data and perform data cleaning; S2. Perform polynomial fitting interpolation for missing trajectory data; use known data to construct a fitting polynomial to fill the missing trajectory data; the fitting polynomial is: Among them, Y m is the predicted dependent variable value, α0 is the intercept term, and α i is the regression coefficient, X i is the value of the ith independent variable in X, and N is the total number of data; S3, normalize the three-dimensional trajectory feature data; the normalization formula for mapping the original data to the interval [0,1] is: Among them, the subscript x is the x-th dimension data in the trajectory vector, the subscript i is the i-th data, f′ is the value after mapping transformation, f is the value before transformation, and f′ xi Represents the i-th original data value f in the x-th dimension xi The value after normalization; F is the set of all vectors before transformation, that is, F = {f1, f2, ..., f i ,…,f N }, F x is the x-th dimension data in the set of all vectors before transformation, and N is the total number of data; S4, processing the small sample data set; segmenting the small sample data set according to the intention, extracting the tracks with the same intention into a new data set while keeping their time order, so that the segmented data set only contains data of a single type of intention; S5. Build the TimeGAN network. The TimeGAN network consists of five parts: generation network, embedding network, recovery network, identification network, and supervision network. The structure of each network is a 24-layer recurrent neural network plus a fully connected layer. S6. Construct a loss function and perform embedding training and supervised training in sequence. S7, construct a joint loss function and perform joint training; S8, build a hidden layer LSTM network to predict the trajectory; S9, construct an output fully connected layer to obtain the three-dimensional coordinate prediction trajectory; In step S6, the output representations of the input X and noise Z after passing through different networks are first given; wherein the noise Z is a random noise uniformly distributed from 0 to 1 with the same shape as the input data X; the output obtained by the input X after the embedding network is represented by H, and the output obtained by H after the restoration network is represented by Indicates that the output of H after the supervision network is Indicates that the output of H after the identification network is Y real Represents; the output of the noise Z after the generation network is expressed as express, The output obtained by the supervised network is express, The output obtained by restoring the network is express, The output obtained by the identification network is denoted by Y fake express, The output obtained by the identification network is denoted by Y efake express; In step S7, first calculate Y fake and Y fake The cross entropy loss for all-1 vectors of the same shape, Next, calculate Y efake and Y efake The cross entropy loss for all-1 vectors of the same shape, in It is used to map any real number to the interval (0,1), N is the total number of samples, Y fake,i Yes fake The predicted value of the i-th sample in efake,i Yes efake The predicted value of the i-th sample in G Gloss =g loss +g lossfr +g lossfre , using the adam optimizer to optimize G Gloss Update Parameters Among them, θ3 is the parameter that needs to be updated, including e g and e s , α represents the learning rate, which is a hyperparameter used to control the step size of each update; It is an unbiased estimate of the gradient, which is used to guide the direction of parameter update; is the variance of the unbiased estimate of the gradient; ∈ is a small positive number used for numerical stability to prevent the denominator from being zero.

2. The track prediction method based on generative adversarial network and long short-term memory network according to claim 1, characterized in that: All parameters of the five networks are stored in a list, and the parameter list of the embedded network is used v Indicates that the parameter list of the recovery network is represented by e r Indicates that the parameter list of the generated network is e g Indicates that the parameter list of the supervision network is represented by e s Indicates that the parameter list of the identification network is represented by e d express.

3. A track prediction method based on generative adversarial network and long short-term memory network according to claim 1 or 2, characterized in that: Calculate the mean squared loss during embedding training: Where N is the total number of data, subscript i is the i-th component in the data, and X i is the i-th component in the input data, for The i-th component in ; the mean square error is used to evaluate the difference between the two, and the adam optimizer is used to optimize e loss Update Parameters Among them, θ1 is the parameter that needs to be updated, including e v and e r , α represents the learning rate, which is a hyperparameter used to control the step size of each update; It is an unbiased estimate of the gradient, which is used to guide the direction of parameter update; is the variance of the unbiased estimate of the gradient; ∈ is a small positive number used for numerical stability to prevent the denominator from being zero.

4. The method for track prediction based on generative adversarial network and long short-term memory network according to claim 3, characterized in that: Compute the mean squared loss during supervised training: Where N is the total number of data, subscript i is the i-th component in the data, and X i is the i-th component in the input data, for The i-th component in g is used; the mean square error of the two is used as the loss of the supervision network, and the adam optimizer is used to optimize g loss Update Parameters Among them, θ2 is the parameter that needs to be updated, including e g and e s , α represents the learning rate, which is a hyperparameter used to control the step size of each update; It is an unbiased estimate of the gradient, which is used to guide the direction of parameter update; is the variance of the unbiased estimate of the gradient; ∈ is a small positive number used for numerical stability to prevent the denominator from being zero.

5. The track prediction method based on generative adversarial network and long short-term memory network according to claim 1, characterized in that: Calculate Y real and Y real The cross entropy loss for all-1 vectors of the same shape, Next, calculate Y fake and Y fake The cross entropy loss of all zero vectors of the same shape, Next, calculate Y efake and Y efake The cross entropy loss of all zero vectors of the same shape, in It is used to map any real number to the interval (0,1), N is the total number of samples, Y fake,i Yes fake The predicted value of the i-th sample in (i.e., the output of the model), Y efake,i Yes efake The predicted value of the i-th sample in G Dloss =g lossrr +g lossff +g lossffe , using the adam optimizer to optimize G Dloss Update Parameters Among them, θ4 is the parameter that needs to be updated, including e d , α represents the learning rate, which is a hyperparameter used to control the step size of each update; It is an unbiased estimate of the gradient, which is used to guide the direction of parameter update; is the variance of the unbiased estimate of the gradient; ∈ is a very small positive number used for numerical stability to prevent the denominator from being zero; in each training batch, θ3 is updated twice and θ4 is updated once.

6. The track prediction method based on generative adversarial network and long short-term memory network according to claim 1, characterized in that: In step S8, the forget gate, update gate and output gate are constructed by the following formula: C f =σ(W f [h t-1 ,x t ]+b f ) C u =σ(W u [h t-1 ,x t ]+b u ) C a =σ(W a [h t-1 ,x t ]+b a ) Among them, Γ f , Γ u and Γ a They represent the forget gate, update gate and output gate respectively, σ is the Sigmoid function, W f , W u and W a is the weight coefficient matrix of the corresponding part, h is the output feature, x is the input feature, subscript t represents the current time, t-1 represents the previous time, b f 、b u and b a is the offset vector of the corresponding part.

7. The method for track prediction based on generative adversarial network and long short-term memory network according to claim 6, characterized in that: The candidate neurons are calculated by the following formula: Among them, W c is the weight coefficient matrix corresponding to this part, b c is the offset vector corresponding to this part; the updated neuron is calculated by the following formula: The output features at the current moment are calculated by the following formula: h t =Γ a *tanh C t 。 8. The method for track prediction based on generative adversarial network and long short-term memory network according to claim 1, characterized in that: In step S9, a fully connected layer is constructed by the following formula: y f =W f ·h t +b f Among them, W f is the weight matrix of the fully connected layer, b f is the bias of the fully connected layer, y f It is the output scalar of the fully connected layer, which contains the predicted trajectory information of one time step. When predicting, it predicts the trajectory points of one time step in one dimension each time, and finally outputs the predicted trajectory of the three-dimensional coordinates corresponding to one time step.

Images