Unmanned aerial vehicle trajectory prediction method and device based on gf-gan and lstm

Abstract

This application relates to a method and apparatus for UAV trajectory prediction based on GF-GAN and LSTM, belonging to the field of UAV trajectory prediction technology. The method includes: acquiring and preprocessing real UAV trajectory time-series data, then inputting it into a GF-GAN generator with a Gaussian filter layer integrated at the end for training; selecting the optimal generator based on the quality evaluation results of the generated trajectory to generate synthetic trajectory data, and fusing it with real trajectory time-series data at different fusion ratios to obtain augmented datasets under different fusion ratios, which are then input into pre-constructed LSTM-based trajectory prediction models for independent training; using each converged trajectory prediction model to perform a UAV trajectory prediction task, and determining the optimal values ​​of key parameters through multi-dimensional comparison to guide the prediction of the UAV's future trajectory under the current task. This method can significantly improve the trajectory prediction accuracy and model generalization ability under small sample conditions.

Description

Technical Field

[0001] This application relates to the field of unmanned aerial vehicle (UAV) trajectory prediction technology, and in particular to a method and apparatus for UAV trajectory prediction based on GF-GAN and LSTM. Background Technology

[0002] Currently, methods for predicting UAV trajectories can be mainly divided into two categories: 1) model-driven methods with dynamic models as the core; and 2) data-driven methods with historical data as the core.

[0003] Model-driven methods rely on precise flight dynamics modeling, requiring accurate knowledge or evaluation of key parameters such as aerodynamic parameters, mass characteristics, and engine thrust. However, the aerodynamic parameters of UAVs in actual flight are difficult to obtain accurately and are subject to significant perturbations, leading to discrepancies between the established mathematical model and the actual physical process, thus limiting prediction accuracy. On the other hand, with the development of artificial intelligence technology, deep learning trajectory prediction methods, represented by Long Short-Term Memory (LSTM) networks, have shown strong advantages. They can automatically learn complex dynamic features from data without requiring precise physical models. However, the performance of deep networks is highly dependent on massive amounts of high-quality training data, which fundamentally contradicts the current situation of high cost of UAV testing and scarcity of available real trajectory data. Direct training based on scarce data easily leads to model overfitting, resulting in a significant decrease in its generalization prediction ability.

[0004] Therefore, although deep learning has great potential for application in UAV trajectory prediction tasks and can effectively capture the temporal dependencies of trajectories, most existing research focuses on improving prediction accuracy from the level of model structure or learning algorithms. In small sample scenarios, these methods are essentially still trying to mine more potential information on a limited basis, failing to fundamentally overcome the core bottleneck of insufficient training data, and making it difficult to achieve simultaneous improvement in UAV trajectory prediction accuracy and generalization ability. Summary of the Invention

[0005] Therefore, it is necessary to provide a method and apparatus for UAV trajectory prediction based on GF-GAN and LSTM to address the above-mentioned technical problems, which can overcome the scarcity of UAV trajectory data and achieve high-precision long-term trajectory prediction.

[0006] A method for predicting the trajectory of a UAV based on GF-GAN and LSTM, the method comprising:

[0007] The real trajectory time series data of the UAV is acquired and preprocessed to construct a Gaussian filter generative adversarial network GF-GAN. The GF-GAN is trained using the preprocessed real trajectory time series data. The generator output of the GF-GAN is equipped with a Gaussian filter layer to perform temporal filtering on the latent space features, thereby enhancing the smoothness of the generated trajectory under small sample conditions and its similarity to the real trajectory time series data.

[0008] The quality of the trajectory data generated by the generator is evaluated according to the preset quality evaluation index. The optimal generator is selected based on the quality evaluation results. The synthetic trajectory data output by the optimal generator and the real trajectory time series data are fused according to different fusion ratios to construct augmented datasets under different fusion ratios. These datasets are then input into the pre-built LSTM-based trajectory prediction model for independent training.

[0009] The UAV trajectory prediction task is performed using each converged trajectory prediction model. By combining the task requirements and comparing the prediction accuracy of each trajectory prediction model under different fusion ratios and prediction step sizes, the optimal fusion ratio and effective prediction step size are determined, and the prediction of the UAV's future trajectory under the current task is guided.

[0010] A drone trajectory prediction device based on GF-GAN and LSTM, the device comprising:

[0011] The trajectory data augmentation module is used to acquire and preprocess real trajectory time-series data of UAVs, construct a Gaussian filter generative adversarial network GF-GAN, and train GF-GAN using the preprocessed real trajectory time-series data. The generator output of GF-GAN is equipped with a Gaussian filter layer to perform temporal filtering on latent space features, thereby enhancing the smoothness of the generated trajectory under small sample conditions and its similarity to real trajectory time-series data.

[0012] The data fusion module is used to evaluate the quality of the trajectory data generated by the generator according to the preset quality evaluation index, select the best generator based on the quality evaluation results, and construct the augmented datasets under different fusion ratios by fusing the synthetic trajectory data output by the best generator and the real trajectory time series data according to different fusion ratios. These datasets are then input into the pre-built LSTM-based trajectory prediction model for independent training.

[0013] The trajectory prediction and key parameter selection module is used to perform UAV trajectory prediction tasks using each trained and converged trajectory prediction model. It combines task requirements and compares the prediction accuracy of each trajectory prediction model under different fusion ratios and different prediction step sizes to determine the optimal fusion ratio and effective prediction step size, and guides the prediction of the UAV's future trajectory under the current task.

[0014] The aforementioned UAV trajectory prediction method and apparatus based on GF-GAN and LSTM significantly improves the quality and usability of generated data by introducing a Gaussian filtering layer into the generator of the Gaussian-filtered generative adversarial network to perform temporal filtering on latent space features. This generates trajectory data with both high fidelity and good smoothness. Furthermore, by proportionally fusing high-quality synthetic trajectory data with original real data, a training dataset with enhanced scale and diversity is constructed, fundamentally solving the bottleneck problem of insufficient data for deep learning model training. Based on this, the LSTM-based trajectory prediction model is trained using the enhanced dataset, enabling it to fully learn the complex spatiotemporal dynamic features of UAV trajectories and achieve high-precision trajectory prediction. This application not only overcomes the dilemma of limited model performance under data scarcity conditions but also provides a basis for the selection of key parameters (such as data fusion ratio and prediction step size) through systematic experimental verification, effectively improving the robustness and practicality of trajectory prediction and overcoming the shortcomings of traditional model-driven methods that rely on accurate physical modeling and conventional data-driven methods that lack generalization ability in scenarios with few samples. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating a UAV trajectory prediction method based on GF-GAN and LSTM in one embodiment.

[0016] Figure 2 This is a schematic diagram of the training process of GF-GAN and LSTM in one embodiment;

[0017] Figure 3 This is a schematic diagram of the Euclidean distance error predicted by the trajectory prediction model under different fusion ratios in one embodiment.

[0018] Figure 4 This is a schematic diagram of the Euclidean distance error predicted by the trajectory prediction model under different prediction step sizes in one embodiment. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0020] In one embodiment, such as Figure 1 and Figure 2 As shown, a method for predicting UAV trajectories based on GF-GAN and LSTM is provided, including the following steps:

[0021] Step 1: Obtain and preprocess the real trajectory time-series data of the UAV, construct a Gaussian filter generative adversarial network GF-GAN, and train the GF-GAN using the preprocessed real trajectory time-series data. The generator output of the GF-GAN is equipped with a Gaussian filter layer to perform temporal filtering on the latent space features, thereby enhancing the smoothness of the generated trajectory under small sample conditions and its similarity to the real trajectory time-series data.

[0022] Step 101: Obtain the actual trajectory time series data of the UAV and perform preprocessing.

[0023] Acquire real-time trajectory data of the UAV, including time and three-dimensional spatial coordinate information. To eliminate numerical instability caused by differences in units and orders of magnitude between different state variables, a min-max normalization algorithm is used to normalize the real-time trajectory data, linearly scaling it to an interval. The normalization calculation formula is:

[0024] ;

[0025] in, The original data values, and These are the minimum and maximum values ​​of the data in this channel, respectively. This is the result after normalization.

[0026] Furthermore, after the model completes trajectory prediction, inverse normalization is performed to restore the trajectory prediction results to the original physical space. The inverse normalization calculation formula is as follows:

[0027] ;

[0028] in, The normalized value of the model's predicted output. These are the actual physical coordinates after restoration.

[0029] It should be understood that this preprocessing process not only helps improve the stability of gradient updates and accelerate convergence during model training, but also enhances the model's robustness to changes in data scale, laying a data foundation for subsequent trajectory prediction tasks.

[0030] Step 102: Construct the Gaussian filter generative adversarial network GF-GAN.

[0031] The core of Generative Adversarial Networks (GANs) lies in the generator. ) and discriminator A minimax game between ( ). Generator The aim is to transform random noise vectors Mapped into the data space to generate trajectory sequences that are sufficiently realistic to be mistaken for real ones. The discriminator As a classifier, it aims to accurately distinguish whether an input sequence comes from a real data distribution. Real trajectory time series data Or from the generator? samples Its objective function can be expressed as:

[0032] ;

[0033] in, Let be the objective function. As the expected value, This indicates the noise distribution.

[0034] In this embodiment, a key improvement was made to the standard GAN structure by designing a Gaussian-filtered generative adversarial network (GF-GAN). Its core innovation lies in integrating a learnable Gaussian filter layer into the generator's output to achieve temporal smoothing optimization of the generated trajectory sequence. This Gaussian filter layer can be used to perform temporal filtering on latent space features, enhancing the smoothness of the generated trajectory under small sample conditions and its similarity to real trajectory time-series data.

[0035] Specifically, the network structure design of GF-GAN is as follows:

[0036] 1. Generator The input is a random noise vector that follows a prior distribution. The noise vector is first expanded in dimension and transformed in feature transfer through a fully connected layer, outputting a preliminary trajectory sequence. .

[0037] To improve the smoothness and physical plausibility of the generated trajectory, this application introduces a learnable Gaussian filter layer at the end of the generator. This Gaussian filter layer... A temporal convolution operation is performed to achieve smoothing filtering. Specifically, this operation is defined in the continuous temporal domain by the convolution integral:

[0038] ;

[0039] in, It is the Gaussian kernel function, and its expression is:

[0040] .

[0041] In practical discrete trajectory sequence processing, the above continuous convolution is equivalent to discrete convolution. The output of the Gaussian filter layer is represented as:

[0042] ;

[0043] in, This represents the discrete trajectory sequence initially output by the generator; This represents a one-dimensional convolution operation; For discrete-time indexing; For discretized Gaussian kernels, specifically representing a kernel of length . The window; These are the parameters for the filter window; Standard deviation of Gaussian kernel A learnable parameter is used for adaptive optimization through network training, enabling GF-GAN to autonomously determine the optimal smoothing intensity best suited to the current trajectory data. The generator's final output is the filtered result. .

[0044] 2. Discriminator The input is real trajectory time series data. Or trajectory data generated by the generator Let them be collectively referred to as the input sequence. ,in For sequence length, For feature dimension, Let be the set of real numbers. The discriminator consists of multiple one-dimensional convolutional layers, used to extract features hierarchically from the input sequence; where, the th... The operation of a convolutional layer is represented as follows:

[0045] ;

[0046] in, and The first The kernel weights and bias terms of a convolutional layer. It is a non-linear activation function. This is the output feature map of the previous layer. .

[0047] To enhance feature robustness and reduce dimensionality, a one-dimensional max-pooling layer is introduced after each one-dimensional convolutional layer, the operation of which is defined as follows:

[0048] ;

[0049] in, For pooling windows in position The covered local area; This is the location index corresponding to the local region. For the first Output feature map of the max pooling layer.

[0050] The resulting feature map after multiple convolution-pooling operations Flattened into a one-dimensional vector The data is then fed into a fully connected layer for final determination.

[0051] ;

[0052] in, and These represent the weights and biases of the fully connected layer. express transpose, It is the Sigmoid activation function. This indicates that the discriminator determines the input sequence. Probabilities derived from the true distribution.

[0053] Through the above design, the GF-GAN constructed in this application not only inherits the powerful distribution fitting ability of GAN, but also explicitly embeds the physical prior of trajectory smoothness into the data generation process by introducing a learnable Gaussian filtering mechanism, thereby guiding the generator to produce high-quality trajectory data that is more in line with the actual dynamic characteristics of UAVs.

[0054] Step 103: Train the GF-GAN using the preprocessed real trajectory time series data.

[0055] The GF-GAN constructed in step 102 is used for adversarial training against preprocessed real trajectory time-series data. The training process aims to solve a minimax game problem by alternately optimizing the parameters of the generator G and discriminator D. The ultimate goal is to enable the generator to produce data with a distribution similar to the real data. Indistinguishable trajectory sequences. GF-GAN uses the standard binary cross-entropy loss function as the optimization objective for training, and the training process employs an iterative approach of alternating between the generator and discriminator:

[0056] First, fixed generator Update the discriminator Sampling random noise from a prior distribution The generator produces the trajectory sequence. At the same time, a batch of data is sampled from the real dataset. Calculate the discriminator loss. The discriminator parameters are updated using backpropagation via gradient descent.

[0057] Then, fix the discriminator. Update generator : Sample new random noise Generate a new trajectory sequence. Calculate the generator loss. The generator parameters are then updated via backpropagation. During this process, the learnable parameters of the Gaussian filter layer at the end of the generator are... It was also optimized in the same way.

[0058] Training continues until the Nash equilibrium point is reached or the maximum number of training iterations are reached. At this point, GF-GAN training is considered complete and can be used for subsequent trajectory data prediction and generation.

[0059] Step 2: Evaluate the quality of the trajectory data generated by the generator according to the preset quality evaluation index, select the optimal generator based on the quality evaluation results, and construct the augmented datasets under different fusion ratios by fusing the synthetic trajectory data output by the optimal generator with the real trajectory time series data according to different fusion ratios. Then, input them into the pre-built LSTM-based trajectory prediction model for independent training.

[0060] Step 201: Evaluate the quality of the trajectory data generated by the generator according to the preset quality evaluation indicators.

[0061] Sample entropy (SampEn) and cosine similarity are introduced as quality evaluation metrics. Sample entropy is used to evaluate the smoothness and intrinsic complexity of the trajectory data generated by the generator, while cosine similarity is used to evaluate the similarity between the generated trajectory data and the real trajectory time series data in terms of overall geometric shape.

[0062] Sample entropy quantifies the regularity of a trajectory by measuring the probability of a pattern repeating in a sequence of trajectory coordinates. A lower sample entropy value indicates a more regular and smoother trajectory, while a higher value indicates a more complex and volatile sequence with poorer smoothness. The formula for calculating the sample entropy of a single trajectory is:

[0063] ;

[0064] In the formula for calculating sample entropy The total number of trajectory points. The length of the trajectory segment used for comparison. This is the tolerance threshold for the distance criterion. It is a length of The distance between the trajectory segments of each point is less than The number of matches, It is a length of The distance between the trajectory segments of each point is less than The number of matches. The sample entropy value of the generated trajectory set is calculated by averaging the sample entropy of each trajectory in the set.

[0065] The cosine similarity metric is used to evaluate the similarity in overall geometric shape between the generated trajectory and the original sample trajectory dataset. Before calculating the cosine similarity between the two datasets, two trajectories are given first. and The calculation process for cosine similarity. For two trajectories... and Each trajectory consists of a series of three-dimensional points: ,in , ,in First, calculate the line segment vector for each trajectory. The line segment vector is , The line segment vector is For each line segment pair ,in The cosine similarity is calculated using the following formula:

[0066] .

[0067] The similarity between two trajectories is the average of the cosine similarities of all line segments, as shown in the following formula:

[0068] .

[0069] Assume the generated and original trajectory sets are respectively and ,in , , and These represent the total number of generated trajectories and the total number of original trajectories, respectively. First, calculate... Each trajectory and The similarity of each trajectory is obtained. ,from Take the middle before The largest similarity value is represented as As shown in the following formula, where This is the batch size during GF-GAN training.

[0070] .

[0071] Therefore, the trajectory and The similarity calculation formula is defined as follows: [Formula omitted for brevity] The average of the maximum similarity:

[0072] .

[0073] Finally, a trajectory set is generated. With the original trajectory set The overall cosine similarity is defined as Each trajectory in and The average similarity is calculated using the following formula:

[0074] .

[0075] Step 202: Based on the quality assessment results, the optimal generator is selected. The synthetic trajectory data output by the optimal generator is then fused with real trajectory time-series data at different fusion ratios to construct augmented datasets with varying fusion ratios. Specifically, the fusion ratio is defined as:

[0076] ;

[0077] in This represents the actual trajectory time series data volume. This refers to the amount of synthetic trajectory data. Adjustments are made to... The value of can be controlled by the system to determine the proportion of generated data in the augmented dataset, thereby constructing multiple augmented datasets with different data compositions for subsequent comparative analysis and model training.

[0078] Step 203: Input the augmented datasets with different fusion ratios into the pre-built LSTM-based trajectory prediction model for independent training.

[0079] LSTM is a special type of recurrent neural network that, through the introduction of sophisticated gating mechanisms, can effectively capture long-term dependencies in time series, making it suitable for prediction tasks with complex temporal dynamics, such as UAV trajectories. The core of an LSTM unit lies in its selective updating and transmission of information through three gating structures: the forget gate, the input gate, and the output gate, as well as the cell state. Its computation process in the next time step is as follows:

[0080] Forgotten Gate: ;

[0081] Input Gate: ;

[0082] Cell status update: ;

[0083] Output gate: ;

[0084] Hidden output: ;

[0085] Among them, the forget gate first determines the current input. and the hidden state of the previous moment Calculate the forgetting weight To determine the cell state at the previous moment. The proportion of information that should be retained in the input. This mechanism gives the model the ability to actively forget redundant historical information; for example, in trajectory prediction, the influence of completed turning actions on the current straight-line motion can be ignored. The input gate further filters new information from the current input, specifically through weights. Controlling candidate cell state The update range, among which, The synergistic effect of the forget gate and the input gate enables the model to dynamically fuse historical and current features, such as integrating the location information of sudden obstacles into motion trajectory prediction. The output gate ultimately regulates the current cell state. The amount of information exposed to the outside world, and the update of the hidden status. , The activation vector of the output gate is used to achieve hierarchical analysis of temporal features, which can capture the local dynamic characteristics of the trajectory and maintain the continuity of the long-term motion trend. This represents element-wise multiplication. It is the sigmoid activation function; , and These are the weight matrices for the forget gate, input gate, and output gate, respectively. , and These are the bias vectors for the forget gate, input gate, and output gate, respectively. For a moment.

[0086] In this application, the input to the LSTM-based trajectory prediction model is the past Historical trajectory sequence at each time step ,in Representing the drone at any time The three-dimensional trajectory coordinates. Output as future. Predicted trajectory sequence at each time step .

[0087] Based on the above input-output definitions, the trajectory prediction model designed in this step consists of an input layer, a stacked multi-layer LSTM network, and a fully connected output layer. The input layer's receiver dimension is... Historical sequence ( (where is the dimension of the state vector). The LSTM network, as the core processing unit, processes the input sequence step by step and finally outputs the hidden state of the last time step. The data is passed to the fully connected output layer, which is responsible for mapping the high-dimensional temporal features extracted by the network back to the physical space. The final output dimension is... The predicted trajectory sequence. In this way, the model can fully learn the trajectory dynamics contained in the augmented dataset and perform long-term, high-precision trajectory prediction.

[0088] During the model training phase, to systematically evaluate the effectiveness of the data augmentation strategy, the model is trained using augmented datasets with different fusion ratios. Each training session is performed independently until the loss function is trained. Loss Below the preset threshold Alternatively, the training can be completed by reaching the maximum number of iterations. The final parameter weights of each model after convergence are retained for subsequent comparative analysis of prediction performance. The augmented dataset with a fusion ratio of 0 is the original dataset.

[0089] Step 3: Utilize each converged trajectory prediction model to perform UAV trajectory prediction task. Based on task requirements, determine the optimal fusion ratio and effective prediction step size by comparing the prediction accuracy of each trajectory prediction model under different fusion ratios and prediction step sizes, and guide the prediction of the UAV's future trajectory under the current task.

[0090] Step 301, Prediction Performance Evaluation. Prediction performance evaluation objectively assesses the effectiveness of this method. Specifically, it evaluates the prediction accuracy of trajectory prediction models trained on different augmented datasets on an independent test set that did not participate in any training process. The evaluation uses two types of metrics: mean absolute error in the three spatial dimensions of longitude, latitude, and altitude, and the combined distance error.

[0091] First, the mean absolute error of the model in the three spatial dimensions of longitude, latitude, and altitude is calculated to accurately reflect the model's prediction accuracy in each spatial direction. The calculation formula is as follows:

[0092] ;

[0093] in Represents longitude, latitude, or altitude. To input the number of test samples, To predict the step size. and They represent the first The test sample in the ... The actual value and the predicted value at each prediction step.

[0094] To evaluate the overall accuracy of trajectory prediction, the Euclidean Distance Error (EDE) is defined as follows: This metric calculates the average Euclidean distance between the predicted and true locations for all test samples across all prediction steps.

[0095] First, regarding the first The test sample in the ... Single-point Euclidean distance error per prediction step for:

[0096] ;

[0097] in, These represent the coordinates of the UAV in different coordinate axes. The formula for calculating the overall Euclidean distance error is:

[0098] ;

[0099] Based on the above evaluation indicators, the prediction performance of the trajectory prediction model under different training sets or different parameter settings can be comprehensively evaluated.

[0100] Step 302, Key Parameter Selection. Based on the quantitative evaluation indicators in Step 301, the following key parameters are selected and analyzed to provide guidance for practical applications.

[0101] The optimal fusion ratio is selected by fixing the prediction step size. Comparing different fusion ratios The prediction errors of each trajectory prediction model are compared, and the optimal fusion ratio is determined based on the comparison results of prediction errors and computational costs. Specifically, the Euclidean distance error predicted by the trajectory prediction model under different fusion ratios is as follows: Figure 3 As shown, by Figure 3 It is evident that as the fusion ratio increases, the introduction of generated data significantly enriches the training set, and the Euclidean distance error of the model prediction continuously decreases. However, the increase in the amount of training data also leads to an increase in computational cost. It is worth noting that after increasing to a certain threshold, the marginal benefit of error reduction significantly decreases. Therefore, determining the optimal fusion ratio requires a trade-off between improving prediction accuracy and controlling computational costs in practical applications, selecting a ratio where the error is reduced to an acceptable level and the computational burden is reasonable as the optimal data fusion ratio. .

[0102] The effective prediction step size is selected by fixing the optimal fusion ratio. Adjust different prediction step sizes The value, combined with the task prediction step size requirements, is analyzed by varying the prediction error. The prediction step size is selected as the effective prediction step size based on the trend of error growth and the inflection point of error growth. . Figure 4 This demonstrates the Euclidean distance error of the trajectory prediction model under different prediction step sizes. Figure 4 It is evident that as the prediction step size increases, the error gradually increases due to the accumulation of uncertainty. The selection of the effective prediction step size should be based on the accuracy tolerance of the specific task: the step size corresponding to the inflection point where the error growth in the error curve begins to steepen should be determined as the effective prediction step size for that task. This ensures reliability while meeting forecasting needs.

[0103] It should be noted that the settings for key parameters are not limited to the values ​​listed in the example. In practical applications, more detailed parameter ranges can be further defined, and the most suitable combination of fusion ratio and prediction step size can be flexibly selected based on the prediction accuracy results and specific task requirements to achieve the best balance between prediction performance and computational cost.

[0104] The aforementioned UAV trajectory prediction method based on GF-GAN and LSTM effectively addresses the fundamental challenge of scarce UAV trajectory data by introducing a learnable Gaussian filter generative adversarial network. This method fully leverages the powerful data distribution fitting capabilities of GANs and applies temporal filtering through a Gaussian filter layer to ensure the quality and reasonableness of the generated data. Furthermore, data augmentation techniques are used to fully unleash the potential of the LSTM model in time-series prediction, ultimately achieving high-precision, long-term trajectory prediction even with limited real-world data. This method, starting from the data source, provides an effective technical path for deep learning applications in small-sample scenarios.

[0105] It should be understood that, although Figure 1 and Figure 2 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order in which these steps are executed, and they can be performed in other orders. Figure 1 and Figure 2 At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0106] In one embodiment, a drone trajectory prediction device based on GF-GAN and LSTM is provided, comprising:

[0107] The trajectory data augmentation module is used to acquire and preprocess real trajectory time-series data of UAVs, construct a Gaussian filter generative adversarial network GF-GAN, and train GF-GAN using the preprocessed real trajectory time-series data. The generator output of GF-GAN is equipped with a Gaussian filter layer to perform temporal filtering on latent space features, thereby enhancing the smoothness of the generated trajectory under small sample conditions and its similarity to real trajectory time-series data.

[0108] The data fusion module is used to evaluate the quality of the trajectory data generated by the generator according to the preset quality evaluation index, select the best generator based on the quality evaluation results, and construct the augmented datasets under different fusion ratios by fusing the synthetic trajectory data output by the best generator and the real trajectory time series data according to different fusion ratios. These datasets are then input into the pre-built LSTM-based trajectory prediction model for independent training.

[0109] The trajectory prediction and key parameter selection module is used to perform UAV trajectory prediction tasks using each trained and converged trajectory prediction model. It combines task requirements and compares the prediction accuracy of each trajectory prediction model under different fusion ratios and different prediction step sizes to determine the optimal fusion ratio and effective prediction step size, and guides the prediction of the UAV's future trajectory under the current task.

[0110] Specific limitations regarding the UAV trajectory prediction device based on GF-GAN and LSTM can be found in the limitations of the UAV trajectory prediction method based on GF-GAN and LSTM mentioned above, and will not be repeated here. Each module in the aforementioned UAV trajectory prediction device based on GF-GAN and LSTM can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0111] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0112] The above embodiments are merely illustrative of several implementation methods of this application, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application.

Claims

1. A method for predicting the trajectory of a UAV based on GF-GAN and LSTM, characterized in that, The method includes: The real trajectory time series data of the UAV is acquired and preprocessed to construct a Gaussian filter generative adversarial network GF-GAN. The GF-GAN is trained using the preprocessed real trajectory time series data. The generator output of the GF-GAN is equipped with a Gaussian filter layer to perform temporal filtering on the latent space features, thereby enhancing the smoothness of the generated trajectory under small sample conditions and its similarity to the real trajectory time series data. The trajectory data generated by the generator is evaluated according to the preset quality evaluation index. The optimal generator is selected based on the quality evaluation results. The synthetic trajectory data output by the optimal generator and the real trajectory time series data are fused according to different fusion ratios to construct augmented datasets under different fusion ratios. These datasets are then input into the pre-built LSTM-based trajectory prediction model for independent training. The UAV trajectory prediction task is performed using each converged trajectory prediction model. The optimal fusion ratio and effective prediction step size are determined by combining the task requirements and comparing the prediction accuracy of each trajectory prediction model under different fusion ratios and different prediction step sizes. This guides the prediction of the UAV's future trajectory under the current task. The GF-GAN consists of a generator and a discriminator with a Gaussian filter layer at the output. The objective function is: ； Among them, generator The aim is to transform random noise vectors Mapped to the data space to generate trajectory data that closely resembles real-world trajectory time-series data; discriminator As a classifier, it aims to accurately distinguish whether an input sequence comes from a real data distribution. Real trajectory time series data Or from the generator? trajectory data , The original data value; Let be the objective function. As the expected value, The noise distribution is represented; the GF-GAN uses the standard binary cross-entropy loss function as the optimization objective for training. The training process is carried out by alternating between the generator and the discriminator until the Nash equilibrium point is reached or the maximum number of training iterations is reached, at which point the GF-GAN training is considered complete. The output of the Gaussian filter layer is represented as follows: ； in, This represents the discrete trajectory sequence initially output by the generator; This represents a one-dimensional convolution operation; For discrete-time indexing; For discretized Gaussian kernels, specifically representing a kernel of length . The window; These are the parameters for the filter window; Standard deviation of Gaussian kernel It is a learnable parameter that is adaptively optimized through network training, enabling GF-GAN to autonomously determine the optimal smoothing intensity that best suits the current trajectory data.

2. The UAV trajectory prediction method based on GF-GAN and LSTM according to claim 1, characterized in that, Acquire and preprocess the real trajectory time-series data of the drone, including: Acquire real-time trajectory data of the UAV, including time and three-dimensional spatial coordinate information; normalize the real-time trajectory data using a min-max normalization algorithm, linearly scaling it to an interval. After completing trajectory prediction, the trajectory prediction results are restored to the original physical space by performing inverse normalization processing.

3. The UAV trajectory prediction method based on GF-GAN and LSTM according to claim 1, characterized in that, The discriminator consists of multiple one-dimensional convolutional layers, used to extract features hierarchically from the input sequence; wherein, the first... The operation of a convolutional layer is represented as follows: ； in, and The first The kernel weights and bias terms of a convolutional layer. It is a non-linear activation function. This is the output feature map of the previous layer. ; The input sequence for the discriminator includes real trajectory time series data. Or trajectory data generated by the generator ; For sequence length, For feature dimension, It is the set of real numbers; To enhance feature robustness and reduce dimensionality, a one-dimensional max-pooling layer is introduced after each one-dimensional convolutional layer, the operation of which is defined as follows: ； in, For pooling windows in position The covered local area; This is the location index corresponding to the local region. For the first Output feature map of the max pooling layer; The resulting feature map after multiple convolution-pooling operations Flattened into a one-dimensional vector The data is then fed into a fully connected layer for final determination. ； in, and These represent the weights and biases of the fully connected layer. express transpose, It is the Sigmoid activation function. This indicates that the discriminator determines the input sequence. Probabilities derived from the true distribution.

4. The UAV trajectory prediction method based on GF-GAN and LSTM according to claim 1, characterized in that, The quality of the trajectory data generated by the generator is evaluated according to preset quality evaluation indicators, including: Sample entropy and cosine similarity are introduced as quality evaluation indicators. Sample entropy is used to evaluate the smoothness and inherent complexity of the trajectory data generated by the generator, and cosine similarity is used to evaluate the similarity between the generated trajectory data and the real trajectory time series data in terms of overall geometric shape.

5. The UAV trajectory prediction method based on GF-GAN and LSTM according to claim 1, characterized in that, The fusion ratio is defined as: ； in This represents the actual trajectory time series data volume. This represents the amount of synthetic trajectory data.

6. The UAV trajectory prediction method based on GF-GAN and LSTM according to claim 1, characterized in that, The LSTM-based trajectory prediction model consists of an input layer, a stacked multi-layer LSTM network, and a fully connected output layer. The input to the model is past data. A historical trajectory sequence at each time step, output as the future... The predicted trajectory sequence at each time step; the LSTM network updates its state through the following gating mechanism: Forgotten Gate: ; Input Gate: ; Cell status update: ; Output gate: ; Hidden state output: ; Among them, the forget gate first determines the current input. Hidden state from the previous moment Calculate the forgetting weight To determine the cell state at the previous moment. The proportion of information that should be retained in the input; the input gate further filters the new information currently being input, specifically through weights. Controlling candidate cell state The update magnitude; the output gate ultimately regulates the current cell state. The amount of information exposed to the outside world, and the update of the hidden status. , is the activation vector of the output gate; This represents element-wise multiplication. It is the sigmoid activation function; , and These are the weight matrices for the forget gate, input gate, and output gate, respectively. , and These are the bias vectors for the forget gate, input gate, and output gate, respectively. For a moment.

7. The UAV trajectory prediction method based on GF-GAN and LSTM according to claim 1, characterized in that, Based on task requirements and by comparing the prediction accuracy of various trajectory prediction models under different fusion ratios and prediction step sizes, the optimal fusion ratio and effective prediction step size are determined, including: The prediction accuracy is evaluated by the average absolute error of the trajectory prediction results in the three spatial dimensions of longitude, latitude, and altitude, as well as the comprehensive distance error. The optimal fusion ratio is selected by fixing the prediction step size. Comparing different fusion ratios The prediction errors of each trajectory prediction model are compared, and the optimal fusion ratio is determined based on the comparison results of prediction errors and computational costs. ; The effective prediction step size is selected by fixing the optimal fusion ratio. Adjust different prediction step sizes The value, combined with the task prediction step size requirements, is analyzed by varying the prediction error. The prediction step size is selected as the effective prediction step size based on the trend of error growth and the inflection point of error growth. .

8. A UAV trajectory prediction device based on GF-GAN and LSTM, characterized in that, The device includes: The trajectory data augmentation module is used to acquire and preprocess the real trajectory time-series data of the UAV, construct a Gaussian filter generative adversarial network GF-GAN, and train the GF-GAN using the preprocessed real trajectory time-series data. The generator output of the GF-GAN is equipped with a Gaussian filter layer to perform temporal filtering on the latent space features, thereby enhancing the smoothness of the generated trajectory under small sample conditions and its similarity to the real trajectory time-series data. The data fusion module is used to evaluate the quality of the trajectory data generated by the generator according to the preset quality evaluation index, select the optimal generator based on the quality evaluation result, and construct the augmented dataset under different fusion ratios by fusing the synthetic trajectory data output by the optimal generator and the real trajectory time series data according to different fusion ratios. The datasets are then input into the pre-built LSTM-based trajectory prediction model for independent training. The trajectory prediction and key parameter selection module is used to perform UAV trajectory prediction tasks using each trained and converged trajectory prediction model. It combines task requirements and compares the prediction accuracy of each trajectory prediction model under different fusion ratios and different prediction step sizes to determine the optimal fusion ratio and effective prediction step size, and guides the prediction of the UAV's future trajectory under the current task. The GF-GAN consists of a generator and a discriminator with a Gaussian filter layer at the output. The objective function is: ； Among them, generator The aim is to transform random noise vectors Mapped to the data space to generate trajectory data that closely resembles real-world trajectory time-series data; discriminator As a classifier, it aims to accurately distinguish whether an input sequence comes from a real data distribution. Real trajectory time series data Or from the generator? trajectory data , The original data value; Let be the objective function. As the expected value, The noise distribution is represented; the GF-GAN uses the standard binary cross-entropy loss function as the optimization objective for training. The training process is carried out by alternating between the generator and the discriminator until the Nash equilibrium point is reached or the maximum number of training iterations is reached, at which point the GF-GAN training is considered complete. The output of the Gaussian filter layer is represented as follows: ； in, This represents the discrete trajectory sequence initially output by the generator; This represents a one-dimensional convolution operation; For discrete-time indexing; For discretized Gaussian kernels, specifically representing a kernel of length . The window; These are the parameters for the filter window; Standard deviation of Gaussian kernel It is a learnable parameter that is adaptively optimized through network training, enabling GF-GAN to autonomously determine the optimal smoothing intensity that best suits the current trajectory data.

Images