China Electronics Technology Group Corporation

Multi-task learning (MTL) is essential for satellite internet systems requiring simultaneous optimization of beam management, interference mitigation, resource allocation, and traffic prediction. However, existing evaluation methods rely predominantly on external performance metrics, neglecting internal dynamics governing task interactions. We propose TDS-Mamba (Task-Aware Decoupled State-Space Model), integrating selective state-space models with task-specific modulation for satellite networks. Our contributions include: (1) Task-Aware Decoupled S6 (TA-DS6) with hypernetwork-generated task-conditioned projection matrices; (2) Shared–Private State Decomposition disentangling cross-task representations from task-specific features; (3) Value-at-Risk (VaR) Gating for risk-sensitive optimization under varying orbital conditions; and (4) an internal diagnostic framework with Task-Specific Entropy and Interference Coefficient metrics. Experiments on LEO satellite constellation benchmarks show consistent improvements over the selected baselines and provide enhanced interpretability of multi-task dynamics via internal diagnostics.

To address the high computational complexity and insufficient real-time performance of traditional multi-radar trajectory planning methods in complex electromagnetic interference environments, this paper proposes an imitation learning-based trajectory planning method for multi-radar systems. This method designs a trajectory policy neural network architecture based on multiple semantic information. It proposes a training data construction method with coverage rate as the optimization objective. Then the trajectory policy neural network is trained by using an imitation learning algorithm with an auxiliary target. Simulation results show that the proposed method achieves an average coverage rate of 93.95%, and improves the single-step decision efficiency by a factor of 6.7 compared with heuristic-based trajectory optimization methods.

Abstract This paper proposes a long short-term memory (LSTM) prediction-enhanced command-filtered backstepping control (CFBC) method to address the low frequency vibration suppression in high-precision air-bearing active vibration isolation systems. A complete nonlinear dynamic model of the vibration isolation system is firstly established which is transformed into multi-input-multi-output (MIMO) high order form. Next, a LSTM state prediction mechanism based on a Gaussian-weighted sliding window is designed, which predicts the weighted trend of future states from historical sequences to improve prediction smoothness and robustness. Building on this, the LSTM prediction output is integrated as a compensation term with CFBC to construct a composite control architecture with leading and filtering capabilities. Through Lyapunov stability analysis, the boundedness of all closed-loop signals and the convergence of tracking errors are proven. Finally, experimental validation on a real air bearing vibration isolation platform shows that, compared to the traditional CFBC method, the proposed strategy significantly reduces tracking errors and effectively improves vibration suppression performance in the key low-frequency band of 1-5 Hz.

Infrared polarization image fusion aims to generate a single image that integrates complementary information from both modalities to enhance scene perception. However, this task is hindered by significant modality gaps, high noise levels in polarization images, and the difficulty of preserving fine details from both sources. To address these challenges, we propose an end-to-end network, the feature-decoupled multi-scale Swin transformer (FDMSFuse). The proposed network uses a multi-scale architecture to capture rich shallow features. Its core component, the mix Swin transformer layer, employs a symmetric shared-key-value attention mechanism for efficient cross-modal interaction. Furthermore, a feature decoupling loss based on a channel-correlation matrix promotes feature complementarity while the DySample module ensures high-quality detail reconstruction. Experiments on the public LDDRS dataset demonstrate that FDMSFuse significantly outperforms nine state-of-the-art methods, ranking first on seven of nine key metrics. Crucially, it improves normalized mutual information (NMI) by nearly 25% over the runner-up, while also achieving top scores for visual fidelity (VIF) and perceived aesthetic quality (NIMA). Qualitative results further confirm its superior performance in noise suppression, texture preservation, and small-target enhancement.

Small object detection in remote sensing imagery remains challenging due to complex backgrounds, frequent occlusions, and dense distributions of objects, which often lead to suboptimal performance with existing models. To address these issues, this paper proposes a novel dynamic element-activated non-semantic sparse attention method for detecting small objects in remote sensing images. First, we introduce a non-semantic sparse attention mechanism that computes self-attention within local patches, enhancing the model’s focus on textures and edges while improving its perception of occluded small objects and local complex variations. Subsequently, a dynamic element-activated cross-layer channel attention mechanism is incorporated to adaptively strengthen cross-layer positional awareness, thereby specifically enhancing the representational capacity of small objects feature against cluttered backgrounds. Finally, a diffusion wavelet convolutional structure is employed to process multi-channel features in parallel, mitigating information loss and capturing critical features of densely distributed small objects under boundary ambiguity. Extensive ablation studies and comparisons with state-of-the-art methods on the VisDrone and AI-TODv2 datasets demonstrate the feasibility and effectiveness of our approach, showing its potential to provide technical support for practical applications in remote sensing small object detection.