Motion State Recognition and Trajectory Prediction of Hypersonic Glide Vehicle Based on Deep Learning

Junbiao Zhang, Jiajun Xiong, Lingzhi Li, Qiushi Xi, Xin Chen, Fan LiIEEE Access 2022 · DOI

TL;DR

Researchers have developed a two-stage deep learning framework that first identifies the specific maneuver state of a Hypersonic Glide Vehicle (HGV) and then uses a hybrid linear/nonlinear model to predict its future trajectory. By combining ConvLSTM networks with self-attention mechanisms, the system achieves over 97% accuracy in state recognition and significantly outperforms existing benchmark models in trajectory prediction.

What Problem Does This Paper Address?

Hypersonic Glide Vehicles (HGVs) are a nightmare for traditional defense systems. Traveling at speeds above Mach 5 in "near space," they don't follow the predictable parabolic arcs of standard ballistic missiles. Instead, they exhibit "jumping" maneuvers in the longitudinal direction and wide-range weaving in the lateral direction.

Current prediction methods—mostly based on parameter extrapolation or simple data-driven models—struggle when an HGV suddenly changes its maneuver state. If a vehicle shifts from a balanced glide to a complex S-turn, traditional models often lag or fail entirely. This paper seeks to bridge that gap by creating a system that is "state-aware."

Methods

The authors propose a modular architecture consisting of two main components:

  1. Motion State Recognition Module: This module uses a ConvLSTM (Convolutional Long Short-Term Memory) network. Unlike standard LSTMs, ConvLSTMs are better at capturing both the temporal (time-based) and spatial (position-based) features of flight data. It classifies the HGV's behavior into one of eight categories based on combinations of longitudinal (balanced vs. jumping) and lateral (none, left-turn, right-turn, weaving) maneuvers.
  2. Trajectory Prediction Module: This is a hybrid "Seq2Seq" (sequence-to-sequence) architecture:
  3. Nonlinear Part: Uses a ConvLSTM-based encoder-decoder with a self-attention mechanism to focus on the most relevant historical data points.
  4. Linear Part: A simple perceptron layer designed to handle the linear characteristics of flight, preventing the complex nonlinear model from over-complicating simple flight segments.

The system uses the output of the recognition module to refine the prediction scheme, allowing it to adapt fluidly as the vehicle changes its flight mode.

Datasets

Because real-world HGV flight data is highly classified and scarce, the authors established a massive HGV Trajectory Library based on high-fidelity physical motion equations.

  • Scale: 2,430 unique trajectories totaling 4.734 million trajectory points.
  • Scope: Covers a wide range of initial conditions (altitudes of 40–70 km, speeds of 3000–6000 m/s) and all eight maneuver modes.
  • Evaluation: The model was tested against simulated radar tracking data to mimic real-world observation conditions.

Key Insights and Ideas

  • The "State-First" Philosophy: The core "aha moment" is that trajectory prediction shouldn't be a monolithic task. By identifying the intent or mode of the vehicle first (e.g., "it's starting a weaving maneuver"), the prediction engine can choose the most appropriate mathematical approach.
  • Hybrid Modeling: The authors recognized that HGV flight isn't always complex. By adding a linear perceptron to the complex Seq2Seq model, they improved robustness, ensuring the model doesn't "hallucinate" complex curves during simple, straight-line glides.
  • Feature Selection: The paper identifies five critical "identification parameters" (height, height change rate, velocity inclination, velocity course angle, and its change rate) that are most sensitive to HGV maneuvers, providing a blueprint for what sensors should prioritize.

Key Results

  • Recognition Accuracy: The ConvLSTM model achieved 97.41% accuracy in identifying the maneuver state, significantly higher than standard RNN (77.51%) or LSTM (82.31%) models.
  • Prediction Precision: In complex scenarios involving both jumping and weaving, the proposed model maintained much lower Root Mean Square Error (RMSE) and Mean Absolute Error (MAE) than state-of-the-art benchmarks like SA-ConvLSTM and EMD_AA.
  • Stability: The model demonstrated a unique ability to maintain stable predictions even at the exact moment an HGV switches from one maneuver type to another.

Limitations and Future Work

While the results are promising, the authors acknowledge several hurdles:

  • Computational Complexity: The model is "time-consuming" to run, which could be a bottleneck for real-time defense applications requiring millisecond responses.
  • Data Dependency: The model's performance is heavily tied to the quality of the training set. If a vehicle performs a maneuver not represented in the library, accuracy may drop.
  • Sensor Noise: While the paper uses filtered radar data, the impact of extreme electronic interference or "plasma sheath" blackouts on the deep learning model remains an area for further study.

Future work will likely focus on optimizing the network for speed and exploring how to make the training sets more robust against "black swan" maneuver types.

Summary updated

Abstract

Hypersonic glide vehicle (HGV) has brought severe challenges to the existing defense system due to its characteristics of high maneuverability, high speed and high precision. Simultaneously, these characteristics also bring great difficulties to trajectory prediction. In this paper, a method for HGV motion state recognition and trajectory prediction based on deep learning is proposed. The proposed method consists of two modules, namely the motion state recognition module and the trajectory prediction module. The motion state recognition module can identify the HGV’s motion state according to state information, and divide it into eight categories. The softmax function is added to the state recognition module to calculate the probability of each motion state. The trajectory prediction module comprises a nonlinear prediction part and a linear prediction part. According to the result of motion state recognition, the appropriate prediction scheme is adopted to better extract the linear and nonlinear characteristics of HGV trajectory, which improves the robustness and prediction accuracy of the proposed method. The experimental results of HGV trajectory prediction show that the proposed method can maintain good stability when the HGV maneuver state changes, and has higher accuracy than the four benchmark methods.
BibTeX
@inproceedings{zhang2022motion,
  author = {Junbiao Zhang and Jiajun Xiong and Lingzhi Li and Qiushi Xi and Xin Chen and Fan Li},
  title = {{Motion State Recognition and Trajectory Prediction of Hypersonic Glide Vehicle Based on Deep Learning}},
  booktitle = {IEEE Access},
  year = {2022},
  doi = {10.1109/access.2022.3150830},
}
Figure