In industrial monitoring, systems typically operate under stable conditions, leading to scarce anomaly samples and thereby limiting the effectiveness of traditional supervised detection methods. However, most existing approaches rely on a single network to process high-dimensional time series, without explicitly capturing fine-grained frequency patterns or complex interdependencies among variables, and often neglect the impact of uncertainty in anomaly detection tasks. To address these challenges, this paper proposes an Uncertainty-aware Frequency-Graph Anomaly Detection (UFGAD) framework for multivariate time series. The framework consists of two complementary branches. First, an uncertainty-aware frequency-Transformer branch is designed to capture refined frequency representations through segment-based multi-angle learning and multi-head attention, while modeling uncertainty from latent representations. Second, a graph deviation network-enhanced graph structure learning branch is introduced to extract inter-feature dependencies using a topological learning mechanism. In addition, a cloud model-based uncertainty quantification strategy is incorporated to measure data-prediction uncertainty within the graph-based framework. Extensive experiments on multiple public datasets demonstrate the strong generalization capability of the proposed method. Furthermore, the framework achieves excellent detection performance in a real-world industrial scenario.
Multi-connected temporal subgraph querying is essential for discovering patterns in real-world systems (e.g., food delivery, financial markets), yet existing methods assume single-edge connections, leading to inaccurate or incomplete results. This work proposes TSMVI, a vertex-index-based approach that enables precise edge binding without repeated vertex or edge traversal. We prove the problem is NP-complete and design a vertex index structure with an edge-relation-code matrix encoding all 13 Allen interval relations. Evaluated on 5 datasets (3 time-series: 310 similar to 8050 nodes; 2 graphs: 451 similar to 21,303 nodes) with query graphs of 3 similar to 9 vertices, TSMVI achieves 100% query accuracy (improving by 6.4% points over baselines) and achieves 1.53 & times; average query speedup over indexed baselines.
Generative Dialogue summarization involves extracting essential information from conversations and distilling it into succinct descriptions. This study introduces a novel multi-task learning framework that combines topic-based data augmentation and external knowledge generation for generative dialogue summarization. Our framework operates by annotating and enriching dialogue data based on topics at the data level, fostering varied expressions across different topics. Additionally, it integrates external knowledge at the entity and paragraph levels to help the model grasp common sense knowledge that is implicit in the context and enhance our model’s understanding of dialogue. Extensive experiments were carried out using two real-world datasets. The results demonstrate that our framework outperforms state-of-the-art models in dialogue summarization tasks, confirming the effectiveness of our proposed methodology.
Aqueous zinc-ion batteries (AZIBs) offer promising safety and affordability, but suffer from dendritic Zn growth and parasitic side reactions at the electrode-electrolyte interface. Herein, we construct a dual-region interfacial modulation framework by molecularly reconfiguring the Helmholtz double layer via trace methyl methacrylate (MMA). Exploiting its amphiphilic and functionally asymmetric architecture, MMA enables a coordinated interfacial reconstruction that disrupts Zn2+ solvation in the outer Helmholtz plane, builds a chemisorbed coordination layer in the inner plane, and modulates local interfacial chemistry with spatial precision. This dualregion regulation collectively suppresses water reactivity, facilitates Zn2+ desolvation, and drives crystallographically preferred deposition along the (101) plane, promoting lateral growth and mitigating dendrite formation. As a result, symmetric Zn||Zn cells exhibit over 4200 h of stable cycling at 1 mA cm-2 and maintain 1100 h of operation at 2 mA cm-2, even at 0 degrees C. Zn||Ti half-cells achieve a Coulombic efficiency of 99.83%, while Zn||NH4V4O10 full cells deliver 93.92% capacity retention after 400 cycles at 2 A g-1, and preserve 85.3% after 300 cycles at 0 degrees C. This work demonstrates a scalable, mechanism-driven electrolyte design paradigm for dendrite-free and high-performance aqueous Zn metal batteries.
Understanding the failure and movement mechanisms of block-type rock slopes is crucial for reliable hazard assessment and mitigation. This study combines physical experiments and numerical simulations to investigate the instability and dynamic behavior of discontinuous slopes using the discontinuous deformation analysis (DDA) method. A laboratory platform equipped with a binocular high-speed imaging system was developed to capture the complete motion process of block sliding, toppling, and collision. Comparative analyses demonstrate that DDA accurately reproduces the experimental kinematics, showing strong agreement in displacement and velocity evolution, and confirming its capability to simulate discontinuous deformation and contact dynamics. The validated DDA model was applied to the Luoma slope along National Highway G318, where a landslide impacted a downstream dangerous rock mass. Simulation results are consistent with field monitoring data, verifying the rationality of the model parameters and physical assumptions. The interaction between the landslide and the dangerous rock mass induced a coupled sliding–toppling failure with high velocity and long runout. When the dangerous rock mass was modeled as intact, it acted as a rigid barrier that effectively attenuated the landslide’s energy and limited its runout distance. The findings provide a scientific basis for quantitative assessment and mitigation of landslide and rockfall hazards in mountainous regions.
