Device-to-device (D2D) is a pivotal technology in the next generation of communication, allowing for direct task offloading between mobile devices (MDs) to improve the efficient utilization of idle resources. This paper proposes a novel algorithm for dynamic task offloading between the active MDs and the idle MDs in a D2D–MEC (mobile edge computing) system by deploying multi-agent deep reinforcement learning (DRL) to minimize the long-term average delay of delay-sensitive tasks under deadline constraints. Our core innovation is a dynamic partitioning scheme for idle and active devices in the D2D–MEC system, accounting for stochastic task arrivals and multi-time-slot task execution, which has been insufficiently explored in the existing literature. We adopt a queue-based system to formulate a dynamic task offloading optimization problem. To address the challenges of large action space and the coupling of actions across time slots, we model the problem as a Markov decision process (MDP) and perform multi-agent DRL through multi-agent proximal policy optimization (MAPPO). We employ a centralized training with decentralized execution (CTDE) framework to enable each MD to make offloading decisions solely based on its local system state. Extensive simulations demonstrate the efficiency and fast convergence of our algorithm. In comparison to the existing sub-optimal results deploying single-agent DRL, our algorithm reduces the average task completion delay by 11.0% and the ratio of dropped tasks by 17.0%. Our proposed algorithm is particularly pertinent to sensor networks, where mobile devices equipped with sensors generate a substantial volume of data that requires timely processing to ensure quality of experience (QoE) and meet the service-level agreements (SLAs) of delay-sensitive applications.

In multi-resolution remote sensing imagery, roads typically exhibit sparse, elongated, and structurally complex morphological characteristics, posing formidable connectivity modeling challenges for semantic segmentation models. Existing approaches predominantly focus on pixel-level accuracy, often neglecting the topological integrity of road networks, which leads to frequent discontinuities and omissions in predicted results. To address this, this paper proposes an end-to-end road extraction framework equipped with multi-receptive field modeling and structural connectivity preservation capabilities. The model incorporates a multi-receptive-field module to capture road patterns across varying spatial scales, a connectivity-aware decoding mechanism to strengthen structural coherence, and a topology-aware loss that explicitly guides the restoration of continuous road networks during training. On the DeepGlobe-Road dataset, TopoRF-Net achieves OA 98.57%, IoU 69.76%, F1-score 82.18%, Precision 85.50%, and Recall 79.12%; on the Massachusetts dataset, TopoRF-Net similarly achieved outstanding results: OA 96.65%, IoU 59.68%, F1-score 74.75%, Precision 77.98%, and Recall 71.77%. These results conclusively demonstrate that the proposed method significantly outperforms existing approaches in both precision and connectivity metrics, whilst exhibiting favorable parameter efficiency and inference performance.

Device to Device (D2D) is a pivotal technology in the next generation of communication, allowing for direct task offloading between mobile devices (MDs) to improve the efficient utilization of idle resources. This paper proposes a novel algorithm for dynamic task offloading between active MDs and idle MDs in the D2D-MEC (Mobile Edge Computing) system by deploying multi-agent deep reinforcement learning (DRL) to minimize the long-term average delay of delay-sensitive tasks under deadline constraints. Our core innovation is a dynamic partitioning scheme for idle and active devices in the D2D-MEC system, accounting for stochastic task arrivals and multi-time-slot task execution, which has been insufficiently addressed in existing literature. We adopt a queue-based system to formulate a dynamic task offloading optimization problem. To address the challenges of large action space and coupling of actions across time slots, we model the problem as a Markov Decision Process (MDP) and perform the multi-agent DRL through Multi-Agent Proximal Policy Optimization (MAPPO). We employ a centralized training with decentralized execution (CTDE) framework to enable each MD to make offloading decisions solely based on its local system state. Extensive simulations demonstrate the efficiency and fast convergence of our algorithm. In comparison with the existing sub-optimal results deploying single-agent DRL, our algorithm reduces the average task completion delay by 11.0% and the ratio of dropped tasks by 17.0%.

Grain warehouses are critical nodes in food storage, circulation, and reserve systems, and their safe operation is relevant to occupational protection, reserve stability, and food-system resilience. However, unsafe warehouse behaviors are still mainly identified through manual patrols and fragmented inspections, making continuous observation, cross-site coordination, and early intervention difficult. To address this problem, this study proposes a grain warehouse safety governance framework with federated visual intelligence. The framework combines a real-time visual perception module, data-local federated learning, and a city-level digital supervision platform. Drawing on the Wenzhou grain reserve application scenario, the study defines twelve visually observable risk categories, including personal protective equipment (PPE) violations, and links artificial intelligence (AI) recognition with multi-source inspection, spatial mapping, warning generation, collaborative verification, work approval, closed-loop rectification, and large-model-assisted analysis. A four-client technical pilot compares isolated local, centralized, Federated Averaging (FedAvg), and Federated Proximal Optimization (FedProx) training across three selected recognition tasks. The task-specific results provide quantitative evidence from held-out warehouse-level validation and an intermittent-client stress test. The best held-out weighted F1-score (F1) values were 0.7712 for crossing over a conveyor, 0.5891 for passing under a conveyor, and 0.8481 for missing five-point safety harness. In the one-small-client-absent stress test, FedProx achieved weighted F1 scores of 0.5701, 0.4324, and 0.8064 for the same tasks, respectively. The framework translates visual detections into warning outputs, management indicators, and accountable safety workflows under data-governance constraints. It positions computer vision (CV) as an automated observation mechanism for responsible AI governance and embeds visual intelligence into a human-confirmed, data-local, and auditable safety-management workflow for sustainable food-storage infrastructure.

Intelligent monitoring in grain-reserve warehouses increasingly combines fixed sensors, UAVs, quadruped robots, human review, and site information systems, yet the resulting operational records are typically stored separately. This fragmentation makes it difficult to reconstruct how a detected problem evolves from observation to execution, evidence, and review, and limits traceable data-driven decision support. To address this issue, this study proposes a Geospatial Operational Event Knowledge Graph (GOEKG) that uses a stable managed-event identity to organize geographic anchors, source provenance, air–ground tasks, evidence, review, and lifecycle states within a property graph. The framework combines source-record normalization and graph mapping with competency queries and operational-semantic constraints for temporal order, spatial continuity, review reachability, lifecycle transitions, and provenance-aware reconciliation. A Panqiao warehouse case is used to instantiate and evaluate the graph. The case instantiated 11 reference-case nodes and 15 typed relations; six competency queries recovered the corresponding paths with temporal and spatial continuity across both robot branches. Controlled mutations of a source-derived record fixture returned UNKNOWN for missing temporal anchors, resolved spatial conflicts by explicit rules, and rejected temporal-order, spatial-linkage, and unauthorized-closure violations. A Neo4j benchmark on synthetic fixtures scaled from 1000 to 50,000 events. Secondary evaluations measured chronological alert-type relation completion and strict temporal fixed-snapshot follow-up candidate ranking. Under the latter protocol, Context-only achieved Macro-F1 0.783 and balanced accuracy 0.836, while Full context achieved average precision 0.952 and prioritized Pending Verification records in the limited-budget queue. The fixed-snapshot analysis supports context-assisted review prioritization under a retrospective event-time protocol. GOEKG therefore provides a traceable event-centered integration layer for intelligent monitoring and review-oriented operational decision support.