16 October 2023 in IEEE Transactions on Aerospace and Electronic Systems
The long-term prediction of hypersonic glide vehicle (HGV) trajectory based on radar tracking data still faces challenges. Existing approaches are hard to provide satisfactory performances in long-term HGV trajectory prediction tasks due to insufficient long-range dependency extraction ability and cumulative errors. To address the above issues, a trajectory prediction model Physics-informed Transformer (PIT) is proposed in this work. The Top $\tau$ -Mean attention mechanism and the generative decoder are designed in this model to extract the trajectories' long-range dependency and reduce the cumulative errors. Furthermore, the physical knowledge is applied at both the input and output ends of the model, which further improves the model's prediction accuracy even in the case of missing a small amount of radar tracking data. A simulated HGV trajectory dataset is established based on the dynamic equations for the PIT trajectory prediction model. The experimental results show that the PIT performs better than other state-of-the-art deep neural network models. The additional experiments on a small-scale dataset and datasets with missing tracking data further verify the excellent performance and robustness of the PIT in these cases.
8 October 2024 in IEEE/CAA Journal of Automatica Sinica
Dear Editor, This letter presents a new transfer learning framework for the deep multi-agent reinforcement learning (DMARL) to reduce the convergence difficulty and training time when applying DMARL to a new scenario [1], [2]. The proposed transfer learning framework includes the design of neural network architecture, curriculum transfer learning (CTL) and strategy distillation. Experimental results demonstrate that our framework enables DMARL models to converge faster while improving the final performance.
3 March 2025 in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
The remote sensing image (RSI) scene classification is currently a popular research topic among many remote sensing tasks. However, RSI scene classification still faces challenges like complex multi-scale key features concentrated in different local regions, large inter-class imbalance, and intra-class variation in insufficient well-labeled RSI samples. To address these challenges, we proposed a novel RSI scene classification method based on an improved vision transformer. This method has better multi-scale feature representation ability due to the improved hierarchical vision transformer structure, in which a feature map fusion layer produces feature maps of different sizes, and a window transformer block with dynamic window shift unit actively shifts to the local region with dense information, flexibly extracting and associating key features with multi-scale in different input regions. Furthermore, we design a curriculum transfer learning framework to alleviate the problems of lacking well-labeled training samples, intra-class variation, and interclass imbalance during the training process of the improved vision transformer. This framework employs a dual-criteria difficulty evaluator to evaluate training samples and provides supplementary supervision to the model by generating training task schedules. Finally, experimental results demonstrate that the proposed vision transformer model achieves rapid convergence and superior performance in RSI scene classification task. The code is publicly available at https://github.com/liuyi-pro/DLTransformer.
22 March 2023 in Neural Computing and Applications
16 December 2025 in IEEE Transactions on Aerospace and Electronic Systems
Providing pilots with explicit and accurate decision-support information through trajectory prediction can help them make prescient maneuvers and gain an advantageous position in close-range air combat. However, the uncertainty and high dynamic of dogfights pose significant challenges for existing flight trajectory prediction methods to achieve satisfactory accuracy, real-time performance, and robustness. To address the above issues, this work proposes the Confrontational Flight Transformer (CFT) for predicting combatant parties' short-term and long-term trajectories in dogfights. A bi-level optimized confrontational encoder is designed to extract spatiotemporal features from trajectories. At the inner level, we proposed the Minimum Information Entropy Sparse (MIES) self-attention mechanism to select the dominant features via the lens of information entropy. At the upper level, we designed a dual-channel encoder block to process the trajectories of both parties separately. A Situation Attention (SA) module is developed to establish a point- wise linkage between trajectories to highlight the relevance in dogfight interactions. We recorded simulated flight trajectories of both parties in dogfights to establish the dataset for training the CFT. Multiple experimental results show that CFT exhibits excellent robustness and real-time performance and outperforms other advanced neural network-based models. We believe the proposed method enables efficient and viable trajectory prediction in air combat and further assists pilots in making better decisions.