Your new experience awaits. Try the new design now and help us make it even better

ORIGINAL RESEARCH article

Front. Neurorobot.

Volume 19 - 2025 | doi: 10.3389/fnbot.2025.1643919

4D Trajectory Lightweight Prediction Algorithm Based on Knowledge Distillation Technique

Provisionally accepted
Weizhen  TangWeizhen TangJie  DaiJie Dai*Zhousheng  HuangZhousheng HuangBoyang  HaoBoyang HaoWeizheng  XieWeizheng Xie
  • Civil Aviation Flight University of China, Guanghan, China

The final, formatted version of the article will be published soon.

To address the shortcomings of current 4D trajectory prediction—namely limited multi‐factor feature extraction and excessive computational cost—this paper proposes a lightweight 4D trajectory forecasting algorithm based on RCBAM–TCN–LSTM knowledge distillation. Our approach innovatively combines the Residual Convolutional Block Attention Module (RCBAM) for powerful spatial feature extraction with the Temporal Convolutional Network–LSTM (TCN–LSTM) for efficient temporal modeling, and employs a teacher–student distillation framework to achieve both model compression and performance enhancement. We use historical ADS B trajectory data from Zhuhai Jinwan Airport, preprocessing raw trajectories via cubic spline interpolation and a uniform‐step sliding window. The teacher network leverages residual structures and channel‐spatial attention in RCBAM to deeply extract high‐dimensional spatiotemporal features, while the student network integrates dilated causal convolutions and a two‐layer LSTM within a TCN–LSTM architecture. In the distillation phase, RCBAM’s predictions serve as soft labels and actual observations as hard labels, jointly guiding student training so that it inherits the teacher’s distribution knowledge while preserving precise real‐world signals. Experimental results in multi‐step prediction scenarios show that the distilled RCBAM–TCN–LSTM model reduces MAE, RMSE, and MAPE by an average of 40%–60% compared to the original RCBAM and TCN–LSTM, and improves R² by 4%–6%. Crucially, this method achieves significant model lightweighting and real‐time deployability, enabling efficient online air‐traffic monitoring and early warning on standard CPUs and embedded devices.

Keywords: 4D Trajectory Prediction, Multi-step prediction, knowledge distillation technique, Teacher-student model, feature extraction

Received: 09 Jun 2025; Accepted: 31 Jul 2025.

Copyright: © 2025 Tang, Dai, Huang, Hao and Xie. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

* Correspondence: Jie Dai, Civil Aviation Flight University of China, Guanghan, China

Disclaimer: All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article or claim that may be made by its manufacturer is not guaranteed or endorsed by the publisher.