Cooperative reconnaissance with multi-UAV swarms poses a high-dimensional, constrained combinatorial optimization problem, where conventional swarm-intelligence methods often suffer from slow convergence and premature stagnation. This paper proposes an Adaptive Lévy Wolf Pack Algorithm (ALWPA) for multi-UAV cooperative reconnaissance task allocation. In ALWPA, the wolf-pack search stages serve only as a structural search framework, while the search behavior is driven by three metaphor-free operators: opposition-based learning for population initialization and renewal, Lévy-flight perturbation for long-range exploration, and differential-evolution recombination for difference-based local refinement. ALWPA was evaluated against ten baseline algorithms on 100 randomly generated instances at each of three problem scales, with the largest comprising 60 UAVs and 100 targets. According to the Friedman and Holm-corrected Wilcoxon tests, ALWPA performed better than every baseline in all 30 pairwise comparisons. It reduced the mean cost by 6–7
While deep learning has significantly advanced environmental sound classification, existing studies mainly focused on closed-set scenarios, with relatively limited exploration of open-set recognition. The scarcity of labeled sound samples further increases the difficulty of learning discriminative representations and reliable class prototypes. To address these challenges, this paper proposes an improved convolutional prototype network for open-set environmental sound classification under limited data conditions. The proposed method contains three complementary components: (1) a time-frequency attention mechanism that emphasizes perceptually salient spectrogram components to enhance discriminative feature learning; (2) a Mahalanobis distance-based complementary prototype loss, which exploits complementary information from misclassified samples to refine prototype representations under limited data conditions; and (3) a learnable Mahalanobis metric for adaptive similarity modeling between feature embeddings and class prototypes. Experiments on environmental sound dataset demonstrate the proposed method consistently improves both known class classification accuracy and unknown class rejection, especially in low-data open-set scenarios.
The reliable operation of industrial systems heavily depends on early and accurate detection of wiring network faults. In this work, we propose a novel diagnostic framework that integrates Time-Domain Reflectometry (TDR) with Temporal Convolutional Networks (TCN). The strength of the proposed TDR-TCN methodology lies in two main aspects: (i) the use of causal and dilated convolutions allows the network to efficiently extract both fine-grained variations and long-range dependencies from TDR signals, and (ii) the architecture demonstrates robustness in managing complex fault scenarios across different wiring configurations, enabling precise fault detection, localization, and characterization. The proposed framework is validated on representative network topologies, achieving outstanding performance: perfect accuracy and macro-averaged sensitivity for a reference Y-shaped network, and a classification accuracy of 99.12
Accurate prediction of spatial interactions is critical for urban planning, transportation management, and regional development. However, existing deep learning approaches often suffer from overfitting, instability, and limited robustness. To address these challenges, this study proposes a multi-stage optimization framework that enhances model performance at both the input and output levels. Specifically, an improved feature selection strategy is employed to identify informative variables and reduce redundancy, while a distribution-aware loss function is introduced to capture the structural characteristics of spatial flows and improve prediction stability. Experimental results on real-world datasets demonstrate that the proposed framework consistently outperforms conventional methods in both accuracy and long-term stability. These findings underscore the importance of task-oriented optimization and provide an effective solution for spatial interaction modeling.
The digital transformation of smart grids has expanded the attack surface for electricity theft and cyber intrusions, yet traditional centralized anomaly detection often compromises user privacy. To address this, we propose FED-XID, a novel federated learning framework that enables data-locality-preserving, distributed anomaly detection. FED-XID introduces three core innovations: (i) an embedded, lightweight XGBoost-based Intrusion Detection System (IDS) for edge nodes; (ii) FedXgbBagging, a novel tree-level aggregation strategy that combines local models through tree concatenation; and (iii) the integration of a hybrid TCN–LSTM module directly into the federated pipeline to perform local data reconstruction, ensuring models learn from high-quality, temporally consistent data. Evaluated on the SGCC dataset (42,372 users), FED-XID achieves an AUC of 93.60
