Consumer-facing low-altitude unmanned aerial vehicle (UAV) services require repeated battery replenishment, while a limited number of charging stations can create uneven charging workloads and reduce fleet availability. This paper studies distributed unbiased extremum seeking (ES) for relaxed UAV charging-station allocation when only delayed local zero-order outputs are available. The charging model combines squared station workloads with a quadratic regularization term, yielding a strongly convex global objective with a unique minimizer. Each UAV receives only its delayed scalar local output and exchanges estimator states over a fixed, connected, undirected graph. Unlike existing unbiased perturbation–demodulation schemes, the proposed phase-modulated Lie-bracket architecture reconstructs the delayed global output over the graph and injects the reconstructed signal directly into the probing phase. The ES baseline is evaluated at the same delayed instant, and exponential scaling removes the steady probing bias. The delay-induced probing phase is retained explicitly in the averaged gradient coefficients. Under local strong convexity, finite-order nonresonance, bounded normalized time-varying measurement-feedback delay, and the stated parameter conditions, local ultimate boundedness is established in scaled coordinates, which yields conditional unbiased convergence of the allocation variables. A six-UAV, two-station simulation with the time-varying measurement-feedback delay demonstrates delayed-global-output reconstruction, allocation convergence, and charging-workload balancing.
Smart-home prosumers increasingly pair rooftop photovoltaics, home batteries, and vehicle-to-grid electric vehicles, and deep reinforcement learning can schedule these assets at near-optimal cost. The policy networks, however, are usually sized by convention and not against the deployment target, so they may carry more parameters than the task requires and exceed the compute and memory budget of a consumer device such as a home energy gateway. This paper addresses that gap with a cloud-edge multi-agent reinforcement learning framework in which the centralized critic is kept large during cloud training to guide compact per-home actors; only those actors ship to the edge and act from local observations alone. The actor depth and per-layer widths are sized automatically by a two-stage Bayesian architecture search driven by a composite deployment objective that jointly penalizes energy cost, degradation, constraint violations, and parameter count. The centralized critic is a heterogeneous graph-attention network that aggregates the joint community state along the electrical neighborhood of the low-voltage feeder, so each home is valued together with the peers to which it is physically coupled. Case studies show that the searched actor uses only about 1.5 k parameters and, against a price-aware rule, lowers the community’s composite objective while increasing local energy trading and reducing peak transformer loading. The framework therefore brings community-scale coordination inside the memory budget of a consumer smart-home gateway.
The rapid growth of consumer internet of things (CIoT) applications increases multimedia traffic and demands secure communication in dynamic low-altitude environments. UAV-assisted networks extend coverage but remain vulnerable to detection, eavesdropping, signal leakage, and CSI aging under mobility. To address these challenges, this paper proposes a closed-loop intelligent covert communication (CLiC) framework for secure UAV-assisted CIoT systems. CLiC combines future-context prediction with dual-timescale hierarchical covert control. A transformer-based prediction module is trained with a supervised forecasting objective. The slow-timescale controller uses a future-context estimate to define the frame-level operating region. The fast-timescale controller then performs feasibility-projected power control, beamforming, and group-wise beyond diagonal reconfigurable intelligent surface (BD-RIS) adaptation at each slot. The UAV-mounted BD-RIS shapes the channel, and group-wise elements control reduces action dimensionality under discrete-phase and coupling constraints. Under 3GPP-compliant channels, CLiC achieves an average secrecy rate of 2.48 bps/Hz and a utility of 2.54. The strongest reactive DRL baseline achieves 0.48 bps/Hz and a utility of 1.49. Covertness is independently evaluated on the realized channel. At the nominal operating point, the average radiometric detection probability remains within the prescribed covert budget, while delivering close to four times more secrecy per unit of detection probability.
Stereo image super-resolution has wide applications in fields such as autonomous driving, consumer electronics, and robot technology, aiming to reconstruct high-resolution stereo images from low-resolution images with effective algorithms. Existing methods face the challenge of extracting high-frequency information in stereo images and reconstructing complex textures with complementary information across stereo image pairs. In order to address above challenge, a Hybrid Wavelet Convolution Attention Network (HWCANet) is proposed to better extract high-frequency information in stereo images with proposed Multi-Level Feature Extraction Block (MLFEB) and fuse the complementary information across left view and right view with proposed Cross-view Wavelet-Enhanced Attention Module (CWEAM). Specifically, different from existing stereo image super-resolution methods applying convolution neural networks or Transformers to extract spatial features, we introduce wavelet transform into our modules to extract frequency features that enrich texture representation and structural details better. Extensive experiments on public benchmarks illustrate that our proposed method is competitive or superior to some state-of-the-art stereo image super-resolution methods.
The increasing penetration of consumer-side solar photovoltaic (PV) systems into utility grids demands reliable real-time monitoring and high-fidelity modeling frameworks capable of operating under dynamically varying field conditions. This paper presents the development and experimental validation of a novel physics-based digital twin (DT) for a single-stage single-phase grid-connected solar PV inverter system. The DT is formulated entirely in the discrete-time domain using the Backward Euler method, driven exclusively by 4 sensor signals. A coupled 2×2 implicit discretization resolves the mutual dependency between the inverter-side inductor and filter capacitor states simultaneously. Unknown system parameters are identified using a novel modified momentum search algorithm (M2SA). Experimental validation on a 6-kW hardware prototype under three operating conditions: varying solar irradiance, grid under voltage, and grid voltage harmonics, demonstrates percentage matching (PM) exceeding 98.5% between physical system (PS) measurements and DT predictions, confirming the robustness, accuracy, and real-time deployability of the DT framework.
