Selected work

Representative Papers

Domain-Adaptive Health Indicator Learning with Degradation-Stage Synchronized Sampling and Cross-Domain Autoencoder

Mar 11, 2026

This work addresses the challenge of distribution mismatch between source and target domains in health indicator modeling under varying operating conditions, particularly caused by misaligned degradation stages and the limited ability of 1D-CNNs to capture long-range dependencies. To tackle this, the authors propose a novel domain adaptation framework that introduces a degradation-stage-synchronized batch sampling strategy to align domains across different operational conditions. Furthermore, they design a cross-domain autoencoder integrating large convolutional kernels with a cross-attention mechanism to effectively learn domain-invariant health representations. Experimental results on the Korean Defense System and XJTU-SY bearing datasets demonstrate that the proposed method outperforms state-of-the-art approaches by an average of 24.1%, significantly enhancing the accuracy and robustness of health indicator construction across diverse operating conditions.

0 citationsRead paper

Exploring Syn-to-Real Domain Adaptation for Military Target Detection

Dec 29, 2025

To address the scarcity of real-world data, high synthetic aperture radar (SAR) acquisition costs, and complex cross-domain variability in military target detection, this paper constructs a high-fidelity RGB synthetic dataset using Unreal Engine—marking the first application of photorealistic synthetic data to domain-adaptive object detection in dynamic military scenarios. We propose a weakly supervised domain adaptation (DA) framework requiring only class-level labels—not pixel- or bounding-box-level annotations—and systematically evaluate mainstream methods (e.g., Faster R-CNN+ADVENT, DCAN) under the Syn→Real transfer setting. Experiments demonstrate that our weakly supervised DA improves mean Average Precision (mAP) by 12.3% over unsupervised DA baselines and significantly outperforms semi-supervised alternatives. These results validate both the practical efficacy and feasibility of lightweight annotation strategies for military vision tasks, highlighting their critical value in resource-constrained operational environments.

0 citationsRead paper

First Attentions Last: Better Exploiting First Attentions for Efficient Transformer Training

Oct 16, 2025

In large-scale Transformer training with tensor parallelism (TP), frequent all-reduce communications between multi-head attention (MHA) and MLP modules per layer impose a severe efficiency bottleneck. Method: We propose FAL, the first architecture that replaces conventional inter-module activation signals with the output of the initial attention layer—enabling a restructured data flow that eliminates all-reduce operations between MHA and MLP within each layer and allows their full parallel execution. We further introduce FAL+, which incorporates normalized attention enhancement and output redirection to improve representational capacity without incurring additional communication overhead. Results: Experiments show FAL achieves up to 44% speedup in multi-GPU training and 1.18× higher single-GPU throughput versus baseline GPT, while attaining lower perplexity. FAL+ further reduces perplexity, demonstrating that communication elimination and model quality improvement are jointly attainable.

0 citationsRead paper

Integrating Symbolic RL Planning into a BDI-based Autonomous UAV Framework: System Integration and SIL Validation

Aug 15, 2025

Autonomous UAV mission planning in dynamic, complex environments demands seamless integration of symbolic planning and reinforcement learning (RL), yet existing approaches struggle with safe, adaptive coordination between rule-based and learned decision-making. Method: This paper proposes AMAD-SRL—a novel framework that for the first time embeds PDDL-based symbolic RL natively into the BDI cognitive architecture, enabling dynamic, safety-governed switching between symbolic and learning-driven planning modes. Guided by domain knowledge modeling, AMAD-SRL supports high-level adaptive behaviors including real-time threat avoidance and target reacquisition. Contribution/Results: Evaluated in a Software-in-the-Loop (SITL) environment, the framework demonstrates stable inter-module coordination and smooth planning-mode transitions. In target acquisition tasks, it reduces flight path length by 75% compared to coverage-based baselines, significantly enhancing mission efficiency and decision robustness under uncertainty.

0 citationsRead paper

Scaling Up without Fading Out: Goal-Aware Sparse GNN for RL-based Generalized Planning

Aug 14, 2025

Existing deep reinforcement learning (DRL)–based generalized planning methods incorporating graph neural networks (GNNs) suffer from severe scalability bottlenecks in large-scale problems due to fully connected graph representations, which induce combinatorial edge explosion, node feature sparsity, and exponential memory growth. Method: We propose a goal-aware sparse GNN architecture that explicitly integrates task-level semantic goals with spatial structural features via spatial-goal-guided local relational modeling and adaptive sparse graph construction. Our approach synergistically combines PDDL symbolic priors with grid-based environmental encoding for efficient, scalable state representation. Contribution/Results: Evaluated on large-scale cooperative UAV mission planning, our method significantly improves policy generalization and task success rates. It is the first to enable end-to-end learned planning in grid environments exceeding 1,000 cells, thereby surpassing the performance limits of conventional dense graph representations.

0 citationsRead paper

Recent publications

Latest Papers

Domain-Adaptive Health Indicator Learning with Degradation-Stage Synchronized Sampling and Cross-Domain Autoencoder

Mar 11, 2026

This work addresses the challenge of distribution mismatch between source and target domains in health indicator modeling under varying operating conditions, particularly caused by misaligned degradation stages and the limited ability of 1D-CNNs to capture long-range dependencies. To tackle this, the authors propose a novel domain adaptation framework that introduces a degradation-stage-synchronized batch sampling strategy to align domains across different operational conditions. Furthermore, they design a cross-domain autoencoder integrating large convolutional kernels with a cross-attention mechanism to effectively learn domain-invariant health representations. Experimental results on the Korean Defense System and XJTU-SY bearing datasets demonstrate that the proposed method outperforms state-of-the-art approaches by an average of 24.1%, significantly enhancing the accuracy and robustness of health indicator construction across diverse operating conditions.

0 citationsRead paper

Exploring Syn-to-Real Domain Adaptation for Military Target Detection

Dec 29, 2025

To address the scarcity of real-world data, high synthetic aperture radar (SAR) acquisition costs, and complex cross-domain variability in military target detection, this paper constructs a high-fidelity RGB synthetic dataset using Unreal Engine—marking the first application of photorealistic synthetic data to domain-adaptive object detection in dynamic military scenarios. We propose a weakly supervised domain adaptation (DA) framework requiring only class-level labels—not pixel- or bounding-box-level annotations—and systematically evaluate mainstream methods (e.g., Faster R-CNN+ADVENT, DCAN) under the Syn→Real transfer setting. Experiments demonstrate that our weakly supervised DA improves mean Average Precision (mAP) by 12.3% over unsupervised DA baselines and significantly outperforms semi-supervised alternatives. These results validate both the practical efficacy and feasibility of lightweight annotation strategies for military vision tasks, highlighting their critical value in resource-constrained operational environments.

0 citationsRead paper

First Attentions Last: Better Exploiting First Attentions for Efficient Transformer Training

Oct 16, 2025

In large-scale Transformer training with tensor parallelism (TP), frequent all-reduce communications between multi-head attention (MHA) and MLP modules per layer impose a severe efficiency bottleneck. Method: We propose FAL, the first architecture that replaces conventional inter-module activation signals with the output of the initial attention layer—enabling a restructured data flow that eliminates all-reduce operations between MHA and MLP within each layer and allows their full parallel execution. We further introduce FAL+, which incorporates normalized attention enhancement and output redirection to improve representational capacity without incurring additional communication overhead. Results: Experiments show FAL achieves up to 44% speedup in multi-GPU training and 1.18× higher single-GPU throughput versus baseline GPT, while attaining lower perplexity. FAL+ further reduces perplexity, demonstrating that communication elimination and model quality improvement are jointly attainable.

0 citationsRead paper

Integrating Symbolic RL Planning into a BDI-based Autonomous UAV Framework: System Integration and SIL Validation

Aug 15, 2025

Autonomous UAV mission planning in dynamic, complex environments demands seamless integration of symbolic planning and reinforcement learning (RL), yet existing approaches struggle with safe, adaptive coordination between rule-based and learned decision-making. Method: This paper proposes AMAD-SRL—a novel framework that for the first time embeds PDDL-based symbolic RL natively into the BDI cognitive architecture, enabling dynamic, safety-governed switching between symbolic and learning-driven planning modes. Guided by domain knowledge modeling, AMAD-SRL supports high-level adaptive behaviors including real-time threat avoidance and target reacquisition. Contribution/Results: Evaluated in a Software-in-the-Loop (SITL) environment, the framework demonstrates stable inter-module coordination and smooth planning-mode transitions. In target acquisition tasks, it reduces flight path length by 75% compared to coverage-based baselines, significantly enhancing mission efficiency and decision robustness under uncertainty.

0 citationsRead paper

Scaling Up without Fading Out: Goal-Aware Sparse GNN for RL-based Generalized Planning

Aug 14, 2025

Existing deep reinforcement learning (DRL)–based generalized planning methods incorporating graph neural networks (GNNs) suffer from severe scalability bottlenecks in large-scale problems due to fully connected graph representations, which induce combinatorial edge explosion, node feature sparsity, and exponential memory growth. Method: We propose a goal-aware sparse GNN architecture that explicitly integrates task-level semantic goals with spatial structural features via spatial-goal-guided local relational modeling and adaptive sparse graph construction. Our approach synergistically combines PDDL symbolic priors with grid-based environmental encoding for efficient, scalable state representation. Contribution/Results: Evaluated on large-scale cooperative UAV mission planning, our method significantly improves policy generalization and task success rates. It is the first to enable end-to-end learned planning in grid environments exceeding 1,000 cells, thereby surpassing the performance limits of conventional dense graph representations.

0 citationsRead paper