Xia Xie

In this study, fresh leaves of the same Camellia sinensis (L.) O. Kuntze cv. Qilan cultivar grown at similar elevations but contrasting latitudes in Wuyishan, Fujian (the 2L-FQ cultivation site), and Baisha, Hainan (the 2L-HQ cultivation site), were systematically compared using environmental characterization, enzyme activity assays, gas chromatography–ion mobility spectrometry (GC–IMS), and untargeted metabolomics. Black teas produced from the two groups of fresh leaves under identical standardized processing conditions were further evaluated using electronic nose and electronic tongue analyses. Correspondingly, 2L-HQ leaves were longer and wider than 2L-FQ leaves, with higher moisture content and water activity. The lower L*, b*, and chroma values of 2L-HQ, together with a ΔE*ab of 6.54, indicated a darker leaf appearance, which was accompanied by higher chlorophyll and carotenoid concentrations. Compared with 2L-FQ, 2L-HQ contained 11.8% more tea polyphenols, 26.4% more soluble polysaccharides, 11.4% more caffeine, and 10.6% more free amino acids. The non-gallated catechins catechin, epicatechin, epigallocatechin, and gallocatechin were significantly enriched in 2L-HQ. DPPH, ABTS, and FRAP antioxidant capacities were also 6.1%, 4.9%, and 2.2% higher in 2L-HQ, respectively. Lipoxygenase, polyphenol oxidase, and peroxidase activities increased by 137.0%, 98.0%, and 93.8%, respectively, whereas β-glucosidase activity remained unchanged. GC–IMS resolved 61 volatile compounds, of which 46 differed significantly between the two origins. 2L-FQ was enriched in several aromatic aldehydes, β-phenyl ethanol, and fruity esters, whereas 2L-HQ contained higher levels of selected alcohols, unsaturated aldehydes, furanic compounds, and lipid-derived esters. Untargeted metabolomics retained 4359 metabolic features, of which 4356 were shared by both groups. After multiple-testing correction, 793 differential metabolic features were identified using VIP > 1, FDR-adjusted p < 0.05, and |log₂FC| > 1, including 420 features with higher abundance and 373 features with lower abundance in 2L-HQ relative to 2L-FQ. These differential features were mainly associated with carbohydrate, amino acid, flavonoid-related, lipid, and cofactor metabolism. Overall, regional cultivation was associated with coordinated differences in leaf morphology, hydration, pigmentation, taste-related composition, antioxidant capacity, oxidative-enzyme potential, volatile fingerprints, and metabolic allocation, providing a data foundation for region-adapted processing of Qilan tea leaves.

To ensure the security and reliability of Internet of Things (IoT) systems, blockchain technology has been widely introduced. The Byzantine Fault Tolerant (BFT) consensus protocol, as the core of blockchain, has thus attracted significant research interest. To enable parallel data processing, a recent line of BFT consensus protocols has adopted Directed Acyclic Graph (DAG) structures, giving rise to DAG-BFT protocols such as BullShark and GradedDAG. These protocols all depend on randomly selected leaders to drive consensus. Yet, in the presence of malicious or slow leaders, mainstream DAG-BFT protocols remain susceptible to significant latency spikes. To mitigate this, recent studies have incorporated reputation-based mechanisms to enhance leader selection in DAG-BFT protocols. However, existing reputation-based protocols still face two fundamental liveness issues, arising from network latency and assumptions of a partially synchronous network. To address these issues, we propose Mako, an asynchronous reputation-based DAG consensus protocol with a dual-path design. Mako employs an optimistic path in which replicas accumulate scores based on their contributions; faster replicas are then preferentially selected as leaders to achieve low latency. Under adverse conditions, Mako transitions to a pessimistic path to guarantee liveness. However, the dual-path execution introduces an additional challenge: ensuring the consistency of blocks committed along the optimistic path and the pessimistic path. To resolve this challenge, we introduce a temporary leader and a rollback mechanism in the pessimistic path.We implement Mako and conduct an extensive experimental evaluation. The results show that Mako significantly outperforms state-of-the-art protocols. In particular, at the 28-node scale, Mako reduces latency by up to 51.8% and improves throughput by up to 26.7% compared to BullShark.

This study proposes a systematic annotation framework to improve the robustness of end-to-end automatic speech recognition (ASR) in a complex low-resource dialect setting, using Hainan Lingao dialect as a case study. The framework consists of three components: semantically complete utterance segmentation instead of fixed-duration clipping; structured annotation at the lexical, sentence, and pragmatic-behavior levels, including explicit tags for dialectal variation, environmental noise, and unintelligible speech as well as rules for handling overlapping speech; and a three-stage quality-assurance workflow with iterative guideline refinement. The framework was implemented in the construction of a Hainan Lingao dialect corpus from 16 speakers and evaluated using 80 h/10 h/10 h training, validation, and test splits under an identical Conformer-based ASR configuration. Compared with a plain-transcription baseline using no special tags and fixed 3 s segmentation, the full specification reduced character error rate (CER) from 8.7% to 7.9%, 24.3% to 18.5%, 19.5% to 15.2%, and 15.2% to 13.1% on clean, noisy, dialogue, and dialect-variation test sets, respectively. The corresponding sentence error rate (SER) decreased from 17.5% to 15.2%, 39.6% to 32.1%, 34.2% to 27.8%, and 28.3% to 24.5%. Ablation experiments further examined the individual contributions of pragmatic-behavior tags, noise tags, semantic segmentation, and dialect-feature annotation. Paired bootstrap testing with 10,000 resamples showed that all baseline-to-full-specification improvements were statistically significant (p < 0.01). These results indicate that systematic annotation can improve ASR robustness in this Lingao low-resource dialect setting, with the largest relative CER reductions observed in the noisy (23.7%) and dialogue (22.1%) scenarios.

The low-altitude economy has experienced significant growth recently, due to the flexibility, affordability, and versatility of low-altitude aircraft. A typical setup uses multiple devices in a distributed system, where the Byzantine Fault Tolerance (BFT) consensus ensures data consistency. However, existing BFT protocols, such as those based on the BKR paradigm with HoneybadgerBFT as a representative, suffer from high communication overhead and latency, limiting scalability and performance. In this paper, we introduce Vega, a novel BFT protocol designed to overcome these challenges. Vega replaces the traditional Reliable Broadcast (RBC) protocol with a more efficient linear Consistent Broadcast (CBC), reducing communication overhead from $O(n^{3})$ to $O(n^{2})$. However, this change introduces a new challenge related to totality, which may impact liveness. Additionally, Vega incorporates a fast path for block agreement, which reduces agreement latency to two communication rounds under optimistic conditions, with a fallback to the original Asynchronous Binary Agreement (ABA) in less favorable cases. However, this introduces another challenge: ensuring consistency between blocks committed via the fast and normal paths. To solve these challenges, we introduce a block retrieval mechanism and a preparation step, ensuring both liveness and consistency. Our experimental results show that Vega significantly outperforms existing protocols, reducing latency by up to 45% and achieving up to 1.8x higher throughput compared to HoneybadgerBFT.

As a promising distributed computing paradigm, edge computing (EC) enhances service quality and reduces service latency by deploying computational and storage resources at the network edge. However, due to edge servers’ geographic distribution and resource constraints, EC is challenged by edge denial-of-service (EDDoS) attacks. Although various approaches have been proposed to mitigate EDDoS attacks, the impact of capacity uncertainties in edge server processing capacity and transmission latency has been largely overlooked. This limits the adaptability and robustness of mitigation strategies in edge computing environments, which could result in overload and paralysis of edge servers. To address this issue, this paper introduces uEDDoS-D, an uncertainty-aware EDDoS mitigation approach based on an improved deep deterministic policy gradient algorithm. uEDDoS-D models the uncertainties of computing capacities and transmission latency of edge servers in EC environments and leverages the collective computational resources of edge servers to mitigate EDDoS attacks without relying on traditional attack detection mechanisms. uEDDoS-D aims to minimize the impact of uncertainties in edge server processing capacity and transmission latency on mitigation decisions while effectively reducing service latency. Experimental results demonstrate that uEDDoS-D significantly reduces service latency, outperforming the state-of-the-art approaches by an average of 16.75%, and enhances system robustness in EDDoS attack scenarios.

Sea state estimation (SSE) is critical to the safety of maritime transport and the reliability of autonomous ships. The frequency of different sea states varies significantly, leading to uneven data distribution. Existing deep learning methods for SSE typically focus on feature extraction, often using simple splicing and fusion, which can result in cross-domain incoherence and degrade model performance. Addressing sea state classification imbalance is often done through distance-based classifiers (e.g., prototype classifiers), but these can be less sensitive to minority classes, and using few prototypes for a class limits the expression of intra-class variations. To overcome these challenges, we propose the Temporal Spectral Cross Fusion Network (TSCFNet), which extracts temporal and spectral features. These are integrated via an innovative temporal spectral cross fusion module to maximize their complementary advantages. Additionally, we introduce a multi-fusion loss function, including temporal, spectral, and fusion losses, to optimize features across different dimensions. This approach improves the performance for minority classes and captures intra-class differences more effectively, solving the problem of category imbalance. Experimental results show that TSCFNet significantly outperforms baseline methods on two imbalanced sea state datasets and multiple multivariate spatio-temporal datasets.

Tables, as an efficient form of structured data representation, are widely applied across domains. However, traditional manual processing methods are inadequate in the big data era, and existing table recognition models, such as SLANet, still face performance limitations. To address these issues, this paper proposes an improved SLANet framework. First, the original H-Swish activation is replaced with the Mish function to enhance feature representation. Second, an end-of-sequence (EOS) termination mechanism is introduced to reduce computational redundancy during inference. Third, a three-phase training strategy is designed to achieve progressive performance improvements. Experimental evaluation on the PubTabNet benchmark demonstrates that the improved SLANet achieves 77.25% accuracy with an average inference time of 774 ms, outperforming the baseline and most mainstream algorithms while retaining lightweight efficiency. The proposed algorithm achieves a TEDS score of 96.67%, significantly surpassing SLANet-based and other state-of-the-art methods. The code will be released upon acceptance.

Wind energy’s rapid global expansion necessitates addressing operational challenges, particularly blade icing, which severely impacts turbine efficiency and safety. Traditional icing detection methods (physical and indirect) often exhibit limited accuracy and real-time performance, especially for low-level icing and in noisy operational environments. While deep learning offers promise, current methods struggle with: modeling the complex spatio-temporal dynamics of icing, adapting to the non-stationarity of feature importance across sensors and time, and maintaining robustness against inherent SCADA data noise. To overcome these deep learning limitations, this paper introduces the Multi-scale Adaptive Robust Aggregation Network (MARA-Net), a novel architecture for robust blade icing detection. MARA-Net’s key contribution lies in its dynamic feature weighting and stochastic pooling mechanism, enabling the model to adaptively prioritize relevant features, learn multi-scale dependencies, and enhance noise resilience. Uniquely combining random and weighted aggregation, MARA-Net effectively addresses the challenges of capturing complex temporal relationships and extracting global context from noisy SCADA data, surpassing the limitations of standard attention mechanisms. Rigorous evaluation using real-world wind farm SCADA data demonstrates MARA-Net’s superior performance in blade icing detection compared to established baselines. Further validation on public datasets confirms MARA-Net’s effectiveness and generalizability. These results underscore MARA-Net’s significant practical potential for enabling more accurate and reliable deep learning-based wind turbine blade icing detection, crucial for improved wind energy operations and safety.