Timegan-based full flight profile time series data generation method

Abstract

This application relates to a method for generating full flight profile time-series data based on TimeGAN. The method divides the original flight data into stages using a sliding window smoothing and threshold determination method, marking takeoff, climb, cruise, descent, and landing stages. After standardizing the multivariate time-series data, a two-layer LSTM TimeGAN model is constructed, and the generator and discriminator are optimized by combining an alternating training strategy and gradient pruning. Altitude correction, velocity limiting, and local cumulative change constraints are applied to the staged generated data to ensure that the generated data conforms to engineering principles. The data from each stage is stitched together by inheriting stage boundary conditions and numerical integration to form a continuous and complete flight profile. This invention significantly improves the structural integrity and physical rationality of the generated data through staged modeling and physical post-processing mechanisms, making the generated data suitable for flight simulation, control verification, and energy management.

Description

Technical Field

[0001] This application relates to the field of aeronautical data simulation and time series generation technology, and in particular to a method for generating full flight profile time series data based on TimeGAN. Background Technology

[0002] With the rapid development of various small and medium-sized aircraft, and the ongoing work on flight mission simulation, control algorithm verification, and energy management strategy development, the demand for high-quality flight profile data is becoming increasingly urgent. However, acquiring real flight data is costly, has limited sample size, and involves privacy concerns, making it difficult to meet the needs of deep learning models for large-scale time-series data. Although Generative Adversarial Networks (GANs) have been widely used in the synthesis of image and speech data, when directly used to generate flight data with complex physical laws and significant stage structures (such as takeoff, climb, cruise, descent, and landing), traditional GANs often fail to guarantee the physical plausibility and engineering usability of the synthesized data. Summary of the Invention

[0003] Therefore, it is necessary to provide a method for generating full flight profile time-series data based on TimeGAN to address the aforementioned technical problems. This method combines phase division, adversarial modeling, and physics post-processing strategies to synthesize complete flight data with engineering constraints and flight laws.

[0004] A method for generating full-flight profile time-series data based on TimeGAN, the method comprising: Obtain the raw flight dataset; the raw flight dataset includes power demand sequences, altitude change sequences, and speed change sequences covering all flight phases.

[0005] The original flight dataset was divided into stages and labeled using a sliding window averaging and threshold determination method, resulting in flight stage label sequences for takeoff, climb, cruise, descent, and landing.

[0006] Based on stage labels, subsets of data for each stage are selected, and the power, climb rate, and acceleration of the corresponding stage are extracted as multivariate time series features. These features are then standardized to obtain standardized time series data.

[0007] A flight data generation model based on TimeGAN is constructed; the flight data generation model includes a generator and a discriminator using a two-layer LSTM structure.

[0008] Standardized time series data were used to conduct adversarial training on the flight data generation model, and alternating optimization and gradient pruning strategies were employed for parameter optimization.

[0009] The trained generator is used to generate data for each flight phase. The generated data for each phase is then post-processed and spliced ​​together according to the flight mission logic to obtain complete flight profile data.

[0010] The aforementioned method for generating full flight profile time-series data based on TimeGAN includes: dividing the original flight data into stages using a sliding window smoothing and threshold determination method, marking takeoff, climb, cruise, descent, and landing stages; constructing a two-layer LSTM TimeGAN model after standardizing the multivariate time-series data, and optimizing the generator and discriminator by combining an alternating training strategy and gradient pruning; applying altitude correction, velocity limiting, and local cumulative change constraints based on the stage-generated data to ensure that the generated data conforms to engineering principles; and stitching together the data from each stage by inheriting stage boundary conditions and numerical integration to form a continuous and complete flight profile. This invention significantly improves the structural integrity and physical rationality of the generated data through stage-based modeling and physical post-processing mechanisms, making the generated data suitable for flight simulation, control verification, and energy management. Attached Figure Description

[0011] Figure 1 This is a flowchart illustrating a method for generating full flight profile time series data based on TimeGAN in one embodiment. Figure 2 This is a diagram illustrating the overall architecture of a TimeGAN-based full-flight profile time-series data generation model in one embodiment. Figure 3 This is a schematic diagram of the generator structure in one embodiment; Figure 4 This is a schematic diagram of the discriminator architecture in another embodiment. Detailed Implementation

[0012] 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.

[0013] This application proposes a TimeGAN-based method for generating full-flight profile time-series data, which introduces a stage division and physical constraint modeling mechanism on top of the TimeGAN structure. Considering the clear differences and physical boundaries between different stages of a flight mission, this method uses stage labels as the modeling entry point, dividing the entire flight process into several structurally stable sub-stages, and training and generating time-series data for each sub-stage separately. Furthermore, a post-processing strategy is introduced, employing methods such as applying velocity change rate constraints, target altitude change rate correction, and velocity amplitude limiting to achieve physical controllability modeling of the synthetic flight data. This method improves the ability of the generated data to fit real flight characteristics, and through the dual effects of stage modeling and physical guidance, significantly improves the continuity, interpretability, and usability of the generated data in engineering tasks.

[0014] In one embodiment, such as Figure 1 As shown, a method for generating full flight profile time series data based on TimeGAN is provided. This method solves the problems of insufficient flight data samples, complex stage characteristics, and strict physical constraints. The method includes the following steps: Step 100: Obtain the raw flight dataset; the raw flight dataset includes power demand sequences, altitude change sequences, and speed change sequences covering all flight phases.

[0015] Specifically, the original flight dataset needs to cover all flight phases, including takeoff, climb, cruise, descent, and landing.

[0016] Step 102: Divide the original flight dataset into stages using the sliding window averaging and threshold determination method, and label them to obtain flight stage label sequences for takeoff, climb, cruise, descent, and landing; Step 104: Filter the subset of data for each stage based on the stage labels, extract the power, climb rate and acceleration of the corresponding stage as multivariate time series features, and perform standardization processing to obtain standardized time series data; Step 106: Construct a flight data generation model based on TimeGAN; the flight data generation model includes a generator and a discriminator using a two-layer LSTM structure.

[0017] Specifically, the generator maps Gaussian noise into normalized time-series data using a two-layer LSTM. The discriminator uses the two-layer LSTM to determine the authenticity of the normalized time-series data generated by the generator.

[0018] The TimeGAN-based flight data generation model framework is as follows: Figure 2 As shown.

[0019] Step 108: Use standardized time series data to perform adversarial training on the flight data generation model, and use alternating optimization and gradient pruning strategies to optimize parameters.

[0020] Step 110: Use the trained generator to generate data for each flight phase. After post-processing the data generated in each phase, stitch them together according to the flight mission logic to obtain complete flight profile data.

[0021] Specifically, flight phase physical rules are applied to the generated data through post-processing methods.

[0022] Altitude correction, velocity limit, and local cumulative change constraints are applied to the phased generated data to ensure that the generated data conforms to engineering principles; by inheriting the phase boundary conditions and stitching together the data from each phase with numerical integration, a continuous and complete flight profile is formed.

[0023] By incorporating flight phase segmentation and physical constraint mechanisms into the training process of generative adversarial networks (GANs), this method overcomes the problem of traditional TimeGANs lacking physical plausibility in generated results under small sample conditions. Through phase-segmented modeling of flight data and imposing speed variation constraints and altitude variation corrections during generation, the physical consistency and engineering usability of the generated data are significantly improved. Compared to general time series generation methods, this method can generate flight profile data that is structurally complete, highly continuous, and physically compliant, making it suitable for scenarios such as flight control simulation and data expansion.

[0024] In the aforementioned method for generating full flight profile time-series data based on TimeGAN, the method divides the original flight data into stages using a sliding window smoothing and threshold determination method, marking takeoff, climb, cruise, descent, and landing stages. After standardizing the multivariate time-series data, a two-layer LSTM TimeGAN model is constructed, and the generator and discriminator are optimized by combining an alternating training strategy and gradient pruning. Altitude correction, velocity limiting, and local cumulative change constraints are applied to the staged generated data to ensure that the generated data conforms to engineering principles. The data from each stage are stitched together by inheriting stage boundary conditions and numerical integration to form a continuous and complete flight profile. This invention significantly improves the structural integrity and physical rationality of the generated data through staged modeling and physical post-processing mechanisms, making the generated data suitable for flight simulation, control verification, and energy management.

[0025] In one embodiment, the original flight data set is segmented by a sliding window average and threshold determination method, including: calculating the first-order difference of the altitude sequence to obtain the altitude change rate sequence; using a mean filter with a preset window size to smooth the altitude change rate sequence to obtain a smoothed altitude change rate sequence; adopting a two-stage determination strategy to first mark the stable altitude cruise stage, and then dividing the remaining stages according to the power demand and change rate threshold; using interpolation filling with adjacent stage labels for the unclassified time periods to obtain a complete stage label sequence.

[0026] Specifically, the original flight data is segmented by a sliding window smoothing and threshold determination method, and the specific steps are as follows 1) Calculate the altitude change rate: Calculate the first-order difference of the altitude sequence to obtain the altitude change rate sequence dh.

[0027] 2) Sliding window smoothing: Use a mean filter with a window size of n1 to smooth the altitude change rate sequence dh to obtain a smoothed altitude change rate sequence dh_smooth.

[0028] 3) Cruise stage marking: Traverse the altitude sequence. If the altitude variance within the window (window half-length is n1 / 2) near a certain time point is less than h1 meters, mark it as the cruise stage (stage label 3).

[0029] 4) Remaining stage determination: Takeoff stage (label 1): Altitude is less than h2 meters and power demand is greater than p1.

[0030] Landing stage (label 5): Altitude is less than h3 meters and power demand is less than p2.

[0031] Climbing stage (label 2): dh_smooth > dh1 m / s Descending stage (label 4): dh_smooth < dh1 m / s 5) Unclassified time period filling: For the unmarked time periods, perform interpolation filling according to adjacent stage labels to ensure the continuity of stage labels.

[0032] In one embodiment, the normalization process in step 104 specifically includes: horizontally splicing the power demand, altitude change rate, speed change rate, and extreme labels to obtain a spliced multi-dimensional time series; normalizing the spliced multi-dimensional time series using Z-score normalization, and retaining the normalization parameters for the inverse transformation of the subsequent generated data.

[0033] In one embodiment, step 108 includes: optimizing the generator and discriminator parameters using the Adam optimizer, and setting the learning rates for the generator and discriminator; performing an alternating update process in each training batch; the alternating update specifically includes: fixing the generator parameters, updating only the discriminator parameters, inputting a mixed sequence of real data and generated data into the discriminator, and minimizing the binary cross-entropy loss function as follows: ; in, The binary cross-entropy loss of the discriminator, For the discriminator to judge real samples x The probability of judging it as true. To generate noise z The generated samples after mapping To provide a discriminator for generating samples The probability of judging it as true; To obtain from real data distribution Samples sampled from the middle, To obtain noise distribution The vector sampled from the middle.

[0034] With discriminator parameters fixed, Gaussian noise is input into the generator, and the generator's loss function is minimized to increase the probability that the generated data is judged as true; the generator loss function is: ; in, This is the generator loss function.

[0035] Gradient clipping is introduced during the generator training phase to limit the L2 norm of the model parameters to no more than 1.0, thus preventing gradient explosion. An attention mechanism is introduced after the two-layer LSTM layer of the discriminator to apply learnable weights to the hidden states at different time steps, resulting in a weighted hidden state... Input the data into the binary classification layer to obtain the discrimination result; set the batch training size to 32 samples and iterate through the standardized stage subset dataset.

[0036] The attention weights are calculated using the following formula: ; in, For attention weights, For the first t Hidden states of the LSTM layer at each time step For trainable weight matrices, Trainable bias vector.

[0037] In one embodiment, such as Figure 3As shown, the generator in step 106 includes: a generator input layer, a two-layer LSTM temporal feature extraction layer, and a fully connected output layer; the generator input layer is used to accept Gaussian noise; the two-layer LSTM temporal feature extraction layer is used to extract features from the Gaussian noise using a sequence-to-sequence temporal modeling method; the fully connected output layer includes a fully connected layer; it is used to map the hidden state of the last time step of the LSTM to the output space consistent with the normalized input dimension, and output normalized temporal data.

[0038] Specifically, the generator network includes: (1) Input layer, which receives Gaussian noise with dimensions [batch_size, 1, input_dim], where input_dim is the standardized feature dimension and batch_size is the batch size.

[0039] (2) A two-layer LSTM temporal feature extraction layer with 64 hidden units in each layer. The feature is extracted by sequence-to-sequence temporal modeling and the activation function is ReLU.

[0040] (3) Fully connected output layer: Map the hidden state of the last time step of LSTM to the output feature space to generate standardized time series data with dimensions (batch_size, sequence_length, output_dim); sequence_length is the sequence length and output_dim is the generator output feature dimension.

[0041] In one embodiment, such as Figure 4 As shown, the discriminator in step 106 includes: a discriminator input layer, a two-layer LSTM temporal feature discrimination layer, an attention mechanism module, a temporal pooling layer, and a fully connected binary classification output layer; the discriminator input layer is used to receive standardized temporal data generated by the generator and real event sequences; the two-layer LSTM temporal feature discrimination layer is used to extract features using a sequence-to-sequence temporal modeling method; the attention mechanism module is used to weight the hidden states at multiple different time steps; the temporal pooling layer is used to extract the hidden state at the last time step; and the fully connected binary classification output layer is used to calculate the true and false probabilities using the Sigmoid activation function.

[0042] Specifically, the discriminator network includes: (1) Input layer, receiving time series data with dimensions [batch_size, sequence_length, output_dim].

[0043] (2) The number of hidden units in the double-layer LSTM temporal feature discrimination layer is symmetrically set with respect to the number of hidden units in the double-layer LSTM temporal feature extraction layer in the generator.

[0044] (3) Temporal pooling layer to extract the hidden state of the last time step; (4) Fully connected binary classification output layer with an output dimension of 1. The binary classification uses the Sigmoid activation function to calculate the true and false probabilities.

[0045] In one embodiment, the post-processing in step 110 includes: performing target mean correction processing on the generated altitude change rate to ensure that the climb, cruise, and descent phases conform to the predetermined altitude change trend; performing mean-removal processing on the generated velocity change rate and applying a sliding window cumulative constraint to limit the cumulative change over any consecutive n time steps to no more than ±a; where n is an integer greater than 0 and a is a real number greater than 0 to limit local velocity abrupt changes; performing phased dynamic limiting of velocity based on phase labels to ensure that the generated velocity curve evolves within a physically reasonable range; and converting the altitude change rate and velocity change rate into a continuous altitude and velocity sequence through numerical integration.

[0046] In one embodiment, the data generated at each stage are post-processed and then stitched together according to the flight mission logic to obtain complete flight profile data. The specific steps of stitching include: using the end altitude and speed of the previous flight stage as the initial boundary conditions for the generation of the next stage to maintain data continuity; integrating the rate of change of speed and the rate of change of altitude at each stage to achieve continuous evolution of altitude and speed under physical constraints; and stitching together the data of each stage in sequence (1. Takeoff → 2. Climb → 3. Cruise → 4. Descent → 5. Landing) based on stage labels to output a standard flight envelope containing timestamps, power requirements, altitude, and speed.

[0047] It should be understood that, although Figure 1 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 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.

[0048] 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.

[0049] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, 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 modifications and improvements all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for generating full-flight profile time-series data based on TimeGAN, characterized in that, The method includes: Obtain the original flight dataset; the original flight dataset includes power demand sequences, altitude change sequences, and speed change sequences covering all flight phases; The original flight dataset was divided into stages and labeled using a sliding window averaging and threshold determination method to obtain flight stage label sequences for takeoff, climb, cruise, descent, and landing. Based on stage labels, the data sets of each stage are filtered out, and the power, ascent rate and acceleration of the corresponding stage are extracted as multivariate time series features. After standardization, standardized time series data are obtained. A flight data generation model based on TimeGAN is constructed; the flight data generation model includes a generator and a discriminator using a two-layer LSTM structure; The flight data generation model is adversarially trained using the standardized time series data, and the parameters are optimized using alternating optimization and gradient pruning strategies. The trained generator is used to generate data for each flight phase. The generated data for each phase is then post-processed and spliced ​​together according to the flight mission logic to obtain complete flight profile data.

2. The method for generating full flight profile time series data based on TimeGAN according to claim 1, characterized in that, The original flight dataset is divided into stages using a sliding window averaging and threshold determination method, including: The first-order difference of the altitude sequence is calculated to obtain the altitude change rate sequence; The height change rate sequence is smoothed by using a mean filter with a preset window size to obtain a smoothed height change rate sequence. A two-stage determination strategy is adopted: first, the highly stable cruise phase is marked, and then the remaining phases are divided according to power demand and rate of change threshold. For unclassified time periods, interpolation between adjacent stage labels is used to fill in the gaps, resulting in a complete sequence of stage labels.

3. The method for generating full flight profile time series data based on TimeGAN according to claim 1, characterized in that, Standardization processes specifically include: By horizontally stitching together power demand, altitude change rate, velocity change rate and extreme labels, a multidimensional time series is obtained. The spliced ​​multidimensional time series is normalized using Z-score standardization, and the standardized parameters are retained for inverse transformation of subsequent data generation.

4. The method for generating full flight profile time series data based on TimeGAN according to claim 1, characterized in that, The flight data generation model is adversarially trained using the standardized time series data, and parameter optimization is performed using an alternating optimization and gradient pruning strategy, including: The Adam optimizer was used to optimize the parameters of the generator and discriminator respectively, and the learning rates of the generator and discriminator were set. In each training batch, an alternating update process is performed; the alternating update specifically includes: With the generator parameters fixed and only the discriminator parameters updated, the discriminator is fed a mixed sequence of real and generated data. The binary cross-entropy loss is minimized as follows: in, The binary cross-entropy loss of the discriminator, For the discriminator to judge real samples x The probability of judging it as true. To generate noise z The generated samples after mapping For the discriminator to generate samples The probability of judging it as true; To obtain from real data distribution Samples sampled from the middle, To obtain noise distribution The vector sampled from the middle; With discriminator parameters fixed, Gaussian noise is input into the generator, and the generator's loss function is minimized to increase the probability that the generated data is judged as true; the generator loss function is: in, The generator loss function; An attention mechanism is introduced after the two-layer LSTM layer of the discriminator to apply learnable weights to the hidden state at different time steps. The weighted hidden state is then input into the binary classification layer to obtain the discrimination result. Set the batch training size to 32 samples and iterate through the standardized stage subset dataset.

5. The method for generating full-flight profile time-series data based on TimeGAN according to claim 1, characterized in that, The generator includes: a generator input layer, a two-layer LSTM temporal feature extraction layer, and a fully connected output layer; The generator input layer is used to accept Gaussian noise; The dual-layer LSTM temporal feature extraction layer is used to extract features from the Gaussian noise using a sequence-to-sequence temporal modeling approach. The fully connected output layer includes a fully connected layer; it is used to map the hidden state of the last time step of the LSTM to an output space consistent with the normalized input dimension, and output normalized time series data.

6. The method for generating full flight profile time series data based on TimeGAN according to claim 1, characterized in that, The discriminator includes: a discriminator input layer, a dual-layer LSTM temporal feature discrimination layer, an attention mechanism module, a temporal pooling layer, and a fully connected binary classification output layer; The discriminator input layer is used to receive standardized time-series data and real event sequences generated by the generator; The dual-layer LSTM temporal feature discrimination layer is used to extract features using a sequence-to-sequence temporal modeling approach; The attention mechanism module is used to weight the hidden states at multiple different time steps; The temporal pooling layer is used to extract the hidden state of the last time step; The fully connected binary classification output layer is used to calculate the true / false probabilities using the Sigmoid activation function.

7. The method for generating full flight profile time series data based on TimeGAN according to claim 1, characterized in that, The post-processing process includes: The generated altitude change rate is corrected for the target mean so that the climb, cruise, and descent phases conform to the predetermined altitude change trend. The generated rate of change is mean-removed and a sliding window cumulative constraint is applied to limit the cumulative change over any consecutive n time steps to no more than ±a; where n is an integer greater than 0 and a is a real number greater than 0. Based on stage labels, the velocity is dynamically limited in stages to ensure that the generated velocity curve evolves within a physically reasonable range; Numerical integration is used to convert the rates of change of altitude and velocity into continuous altitude and velocity sequences.

8. The method for generating full flight profile time series data based on TimeGAN according to claim 1, characterized in that, After post-processing the data generated at each stage, the data is stitched together according to the flight mission logic to obtain complete flight profile data. The specific stitching steps include: The terminal altitude and velocity of the previous flight phase are used as the initial boundary conditions for the generation of the next phase to maintain data continuity; By integrating the rate of change of velocity and the rate of change of altitude at each stage, the continuous evolution of altitude and velocity is achieved under the condition of satisfying physical constraints. The data from each stage are concatenated and combined sequentially based on stage labels to output a standard flight envelope containing timestamps, power requirements, altitude, and speed.

Images