Air quality monitoring has been revolutionized by the use of low-cost sensors (LCS) mounted on the Internet of Things (IoT) nodes. These LCS technologies enable finer monitoring resolutions, e.g., measurements at street level or in specific locations of interest. When combined with government-ruled reference stations, these nodes establish heterogeneous measurement networks that allow for finer spatial resolution. However, a major problem with using LCS is their long-term data quality, as both sensors and calibration models tend to suffer from drift and concept drift, respectively. Recently, machine learning techniques, in conjunction with periodic and opportunistic recalibration and remote correction, have been proven to provide enhanced data quality performance. In this paper, we propose a hybrid mechanism that leverages remote correction based on virtual sensors alongside periodic recalibration to improve long-term data quality while reducing the operational costs associated with node co-location for recalibration. The results demonstrate that the proposed mechanism enhances the viability of LCS for long-term monitoring deployments by an improvement of up to 17%, compared to traditional recalibration approaches, while reducing operational costs.
This paper develops a new spectrum sharing scheme for Vehicle-to-Everything (V2X) systems, termed multi-agent reinforcement learning graph emergent communication (MARLGEC). Although various decentralized multi-agent reinforcement learning (MARL) based spectrum allocation approaches have been introduced to cope with the rapidly time-varying channels in highly mobile vehicular environments, their limited global awareness often leads to poor generalization and pronounced learning non-stationarity. To address these issues, each vehicle in the proposed framework is modeled as an autonomous agent equipped with a learnable emergent communication module, which enhances global perception for decentralized agents and alleviates the non-stationarity caused by partial local observations. Furthermore, a graph-attention-driven communication topology is employed to prune redundant message exchanges, thereby reducing communication overhead. In this manner, the proposed scheme improves the reliability of V2V payload transmission under dynamic traffic and channel conditions. Simulation results validate the effectiveness of the proposed method, demonstrating its adaptability to dynamic environments and outperforming existing MARL-based approaches.
Vehicular networks are increasingly exposed to both in-vehicle bus attacks and external communication threats, creating a need for intrusion detection that combines high classification accuracy with low deployment overhead. Existing lightweight models, however, often have difficulty preserving local structural irregularities while simultaneously modeling higher-level semantic dependencies. This paper proposes a dual-branch heterogeneous spatial-semantic fusion framework for image-encoded vehicular traffic. A split-before-fit protocol prevents held-out records from influencing learned preprocessing parameters. Before imputation, Min–Max scaling, autoencoder fitting, Isolation Forest configuration, quantile recalibration, or image construction, raw records are partitioned into disjoint training and held-out test sets. Each fitted transformation is frozen before it is applied to the test records. Images are then constructed separately within each split from fixed, non-sliding, non-overlapping windows of 24 records; labels are assigned only after a window has been formed by its majority record label. Thus, no raw record contributes to images in both splits. The spatial branch models local patterns induced by the deterministic record–feature layout, whereas the semantic branch uses MBConv-based stages and attention to model compact cross-channel dependencies. Balanced feature alignment, residual fusion mapping, cross-branch calibration, and collaborative re-weighting progressively integrate the two representations. The complete model contains 1.012 M parameters and requires 0.364 GMACs (0.727 GFLOPs) for a 224×224×3 input. Under the adopted split-before-fit record-level hold-out protocol, the model achieves 99.064% Accuracy and 99.064% F1-score on Car-Hacking_5%, and 98.234% Accuracy and 98.235% F1-score on CICIDS-2018. Class-wise performance and confusion-matrix analyses further identify the remaining error patterns. These are closed-set, within-dataset results under a randomized record-level hold-out protocol. They demonstrate neither temporal separation nor cross-session, cross-vehicle, cross-dataset, or unknown-attack generalization.
With the rapid development of computation-intensive and delay-sensitive mobile applications, edge networks face growing demand to provide timely computing services for workloads that are widely distributed and change dynamically over time. In areas where terrestrial edge infrastructure is unavailable, overloaded, or unable to follow shifting service demand, unmanned aerial vehicles (UAVs) equipped with mobile edge computing (MEC) servers can provide computing resources through flexible deployment and controllable mobility. Nevertheless, limited communication, computing, and onboard energy resources complicate the timely processing of tasks with heterogeneous delay tolerances and priorities. This paper investigates joint task offloading and trajectory optimization in a multi-UAV-assisted MEC system. A priority-weighted timeliness metric evaluates task completion relative to individual delay requirements, while the service model accounts for time-division multiple access contention and UAV-side queueing. To coordinate decisions with different time scales and variable structures, we develop a hierarchical Lagrangian learning-and-optimization framework. At the upper level, multi-agent twin delayed deep deterministic policy gradient (MATD3) learns continuous multi-UAV trajectory policies. At the lower level, majorization-minimization block coordinate descent (MM-BCD) optimizes task association, offloading ratios, and service sequences, while Lagrangian multipliers account for the long-term energy budget. Simulation results indicate that the proposed framework generally achieves lower timeliness loss and higher task completion rates than the considered baselines, with more evident gains under heavy network loads.
