Existing studies on multi-UAV inspection methods largely overlook path overlap avoidance among UAVs and differentiated quality control of inspection tasks. As a result, such systems are unable to ensure both cooperative efficiency and the inspection quality of critical areas in practical operations. To address these limitations, this paper proposes a framework that integrates a path overlap avoidance mechanism with a task importance grading strategy. The framework adopts a hybrid decision-making architecture combining centralized task allocation with distributed reinforcement learning control. Specifically, a central controller performs global task assignment based on overall task importance, real-time UAV status, and available bandwidth. Meanwhile, individual Unmanned Aerial Vehicles (UAVs) employ reinforcement learning to autonomously conduct local trajectory planning and imaging decisions. Through the path overlap avoidance mechanism, redundant inspection of the same area is effectively prevented, while bandwidth is dynamically requested or released according to task importance. Simulation results demonstrate that, compared with existing approaches, the proposed method improves inspection efficiency by approximately 24% and enhances overall inspection quality by 20%.

Show more

To address the issue of Quality of Service (QoS) degradation for mission-critical tasks in Low Earth Orbit (LEO) satellite edge computing under extreme congestion, this paper proposes an intelligent computation offloading architecture based on graph attention imitation learning. The proposed method models the satellite network as a time-varying graph to extract global topological features and innovatively introduces a QoS-aware action masking mechanism to forcibly preclude suboptimal decisions through strict physical constraints. Simulation results demonstrate that the proposed algorithm significantly reduces both the network-wide average latency and the 95th percentile tail latency. Furthermore, it achieves the highest success rate for high-priority tasks under heavy-load conditions. Ultimately, this architecture effectively overcomes the "herd effect" among local nodes, realizing superior global load balancing and absolute QoS guarantees for critical tasks within dynamic satellite networks.

Show more

Multimodal Sentiment Analysis (MSA) seeks to predict a speaker's sentiment orientation by comprehensively utilizing modalities such as text, vision, and audio. As deep learning and cross-modal fusion technologies evolve, key challenges include alleviating heterogeneity across modality feature spaces, avoiding bias from fixed main-modal fusion strategies, and enhancing model adaptability to dynamic changes in modality contribution across different samples. To address these issues, this paper proposes a multimodal sentiment analysis framework based on adaptive modality selection and contrastive learning alignment, named Adaptive Modality Selection and Guided Fusion Network (AMSGFN). The framework first employs a cross-modal contrastive learning alignment mechanism to map text, vision, and audio features into a shared semantic space, mitigating semantic discrepancies among heterogeneous modalities. A lightweight modality scoring module then evaluates the discriminability and reliability of each modality for the current sample, adaptively identifying the dominant modality. Building on this, a dominant modality-guided fusion mechanism selectively integrates supplementary information from auxiliary modalities around the dominant modality, highlighting key emotional semantics while suppressing noise and redundant information. Experimental results demonstrate that the proposed method achieves superior performance compared to existing approaches across multiple public datasets, confirming the effectiveness and robustness of the framework in multimodal sentiment analysis.

Show more

The rapid proliferation of Internet of Things (IoT) devices necessitates the deployment of lightweight Network Intrusion Detection Systems (NIDS) at the network edge. Knowledge Distillation (KD) has emerged as a prevailing paradigm to compress cumbersome deep learning models for resource-constrained environments. However, existing KD approaches—whether employing full-representation transfer or static gradient masking—indiscriminately distill both generalized attack signatures and dataset-specific environment noise. This rigid feature entanglement leads to severe negative transfer, exacerbates the catastrophic forgetting of long-tail attack categories, and dramatically degrades cross-environment generalization. To overcome these critical limitations, we propose Disentangled Representation Distillation (DRD), a novel framework that fundamentally alters the nature of transferred knowledge. DRD compels the high-capacity teacher model to decouple high-dimensional network traffic representations into two orthogonal latent sub-spaces: pure Attack Semantics and Environment Background. During the distillation phase, dataset-specific background noise is explicitly discarded, allowing the student model to exclusively inherit purified attack logic. This mechanism circumvents the computational overhead of dynamic masking while equipping the student with inherent robustness against domain shifts. Comprehensive experiments across benchmark datasets (e.g., UNSW-NB15 and CICIDS2017) demonstrate that DRD significantly improves the detection of rare, long-tail attacks, achieves unprecedented cross-dataset generalization, and maintains an ultra-lightweight footprint suitable for edge deployment.

Show more