Electric Power University

Radio-frequency (RF) sensing enables the surveillance of unmanned aerial vehicles (UAVs) beyond the visual line of sight, but recognition reliability can degrade when propagation conditions, interference, and target identity vary. We present an empirical study of learning-based model families that operate on raw in-phase and quadrature (I/Q) samples for UAV RF individual recognition on DroneRFb-DIR, a public dataset containing six commercial drone types, three physical individuals per type, one background class, urban line-of-sight (LOS) and non-line-of-sight (NLOS) acquisition, and six held-out individuals that appear only in the test set. The study compares a raw-I/Q patch Transformer with a monolithic raw-I/Q Transformer, a burst-timing gated recurrent unit (GRU) model, a magnitude-spectrogram model based on the Audio Spectrogram Transformer (AST), and a phase-ablated raw-I/Q variant under a robustness-oriented protocol involving propagation variation, seen-to-test-only identity shift, background rejection, and timing perturbation as a shortcut diagnostic. The raw-I/Q patch Transformer achieves the strongest aggregate performance, with a macro-averaged F1-score (Macro-F1) of 96.84 ± 0.54%, a balanced accuracy of 97.12 ± 0.41%, and an area under the receiver operating characteristic curve (AUROC) of 0.982 ± 0.006 for seen-versus-test-only separation. It also retains the highest performance in the harder evaluation regime (type-level Macro-F1), with the smallest easy-to-hard degradation of 2.56 percentage points. Fine-grained type-level, individual-level, and background-level analyses further document the performance of the evaluated raw-I/Q configuration without isolating the signal mechanisms responsible for the observed differences. Under the shared 1024-sample protocol, the proposed model has the highest mean closed-set scores, identity-shift AUROC and shifted-condition F1, and the smallest timing degradation, in comparison with five recent task-specific adaptations. The domain-adversarial adaptation has the smallest propagation degradation.

Seafood spoilage produces volatile basic compounds that can be detected from package headspace, but practical colorimetric freshness labels must remain stable and quantitative under the high-humidity conditions typical of seafood packaging. The purpose of this study was to develop a humidity-tolerant headspace freshness indicator and to evaluate whether its color response could be quantitatively translated into seafood shelf-life decisions. Pitaya-peel betacyanin was immobilized in a pectin-chitosan polyelectrolyte complex film to improve pigment retention, reduce humidity-induced baseline drift, and maintain sensitivity to volatile bases. The resulting label showed a monotonic color response to NH3 over the tested concentration range and retained a strong signal under near-saturated relative humidity. Compared with single-polymer films, the pectin-chitosan matrix reduced pigment leaching and improved color stability in wet headspace conditions. A two-step calibration workflow was established by converting color difference, Delta E, first to predicted headspace NH3 and then to total volatile basic nitrogen, TVB-N. In packaged whiteleg shrimp stored under refrigerated and retail-mimic temperature-abuse conditions, the label response closely followed the increases in TVB-N, total viable count, and pH during spoilage. Across 45 package-time points, TVB-N prediction achieved RMSE values of 2.13-2.24 mg N/100 g and R2 values of 0.947-0.960. Freshness classification into Fresh, Warning, and Spoiled categories reached an overall accuracy of 0.867 with Cohen's kappa = 0.792. These results indicate that the PEC-betacyanin label can provide a humidity-resilient and quantitative tool for non-destructive seafood freshness monitoring and shelf-life decision support.

This paper investigates the boundedness properties of a convolution operator associated with the fractional Fourier transform (often abbreviated as FrFT) of order α . The operator under study incorporates chirp weight functions γ _1,2(x)=e^± i(x-a(α )x^2) and serves as a natural generalization of classical convolution structures within the fractional domain, which was first introduced in Wirel. Pers. Commun. 92, 623–637, (2017). First, we establish a Young-type inequality for this operator, proving that it maps L_p(ℝ)× L_q(ℝ) into L_r(ℝ) for p,q,r>1 satisfying 1/p+1/q=1+1/r , with an explicit constant depending solely on α . Second, we prove a Hausdorff–Young type inequality which guarantees boundedness into the conjugate space L_s_1(ℝ) whenever 1≤ p,q,s≤ 2 and 1/p_1 +1/q_1=1/s , thereby extending the range of exponents to the dual setting. Third, we derive a Saitoh-type weighted inequality valid for all p>1 (including p=2 ), providing L_p -boundedness with respect to suitable weight functions. The proofs rely on the Riesz-Thorin interpolation theorem, Hölder’s inequality, and Plancherel-type identities for the FrFT. These results unify and substantially extend classical convolution estimates to the fractional Fourier framework.

Full-waveform inversion (FWI) is a method that utilizes seismic data to invert the physical parameters of subsurface media by minimizing the difference between simulated and observed waveforms. Due to its ill-posed nature, FWI is susceptible to getting trapped in local minima. Consequently, various research efforts have attempted to combine neural networks with FWI to stabilize the inversion process. This study presents a bidirectional physics-constrained FWI framework that leverages transfer learning by pre-training on simple initial models and utilizing the results. Additionally, it employs FWI gradients to co-optimize both the neural network and the adaptive residual learning module under bidirectional physics constraints. By eliminating the reliance on a large amount of manually constructed synthetic data sets, the proposed training strategy addresses the challenge of data dependence. Furthermore, through the joint optimization strategy guided by bidirectional constraints, the neural network is able to focus on integrating physically informed prior knowledge into global stratigraphic representations, while the adaptive residual learning module specializes in learning residual mappings from the network's output, thereby capturing subtle interlayer velocity variations in local geological structures. Evaluating the method on two benchmark models under various conditions, including absent low-frequency data, noise interference, non-uniform receiver configurations and differing initial models, along with corresponding ablation experiments, consistently demonstrates the superiority of the proposed approach.

In tropical megacities undergoing rapid industrialization, comprehensive assessments of seasonal variability in heavy metal(loid)s sources and ecological risks in river-canal sediments remain limited, particularly in Southeast Asian urban systems. Urban sediments function as both secondary sinks and potential sources of heavy metal(loid)s, yet integrated evaluations linking geochemistry, source apportionment, and predictive risk modeling are still unexplored. A total of 144 sediment samples were collected during the dry and wet seasons of 2022. Here, we applied an integrated methodological framework combining geochemical characterization, sequential extraction for metal speciation, non-negative matrix factorization (NMF) for source apportionment, GIS-based risk mapping, and machine-learning prediction using Random Forest and XGBoost, with SHAP interpretation. Results demonstrated pronounced spatial and seasonal heterogeneity. Metal burdens were dominated by Fe, Cr, Cu, Zn, and As, primarily derived from crustal/soil inputs and traffic-related urban emissions. Sequential extraction indicated that Cr and Fe were largely associated with residual fractions, whereas Zn, Cu, and As were enriched in redox-sensitive phases. NMF resolved five source profiles, and risk mapping consistently identified high-risk hotspots (levels 4-5) in the Nhieu Loc-Thi Nghe canal, with additional concerns in Ben Nghe canal. Among predictive models, XGBoost exhibited superior performance, and SHAP analysis identified Cr, Fe, and As as the dominant risk drivers. Overall, this study establishes a data-driven, interpretable framework that supports evidence-based prioritization of remediation zones and strategic urban sediment management. Future work integrating metal bioavailability, ecotoxicological responses, and hydrodynamic processes would further refine risk prediction and sustainable intervention planning.