Search Results (1,594)

Remaining Useful Life (RUL) prediction for aircraft engines is a critical task in prognostics and health management, aiming to extract degradation information from multi-sensor operational data to support maintenance scheduling and operational risk management. Existing hybrid temporal models have explored various combinations of convolutional networks, temporal convolutions, recurrent networks, and attention mechanisms; however, there remains room for further coordination regarding the positional roles of different modules within fixed windows, bidirectional context organization, and training objectives. To address this, we propose a Multi-scale Bidirectional Query-Guided Gated Network, MBQG-Net. MBQG-Net employs one-dimensional convolution and parallel bidirectional dilated convolutions to extract local and multi-range temporal features, and utilizes multi-head Query–Value temporal attention to generate window-level context-enhanced representations prior to recurrent state aggregation. The sequence enhanced by Query-based temporal weighting and Value aggregation is then fed into a stacked bidirectional GRU for subsequent bidirectional gated state aggregation. During training, an asymmetric weighted mean squared error is adopted, assigning higher weights to RUL overestimation errors. MBQG-Net was evaluated on all four official test sets (FD001–FD004) of the NASA C-MAPSS dataset under a unified experimental protocol. It achieved the strongest overall regression performance on FD001 and FD003 and maintained competitive overall conventional regression performance on the multi-condition FD002 and FD004 subsets. The model obtained the lowest NASA Score among the compared methods on all four subsets, although the difference on FD002 was marginal. On FD004, MBQG-Net achieved a NASA Score of 930.9088, representing a 28.27% reduction relative to the second-best result. Experimental results demonstrate that MBQG-Net maintains a favorable balance between conventional regression accuracy and direction-sensitive error control across different operating-condition and fault-mode settings, offering a structurally clear hybrid temporal modeling solution for multi-sensor aircraft engine RUL prediction. Full article

Background: The reliability of generative artificial intelligence for Chinese medical information remains uncertain. This study evaluates the concordance and linguistic performance of three generative artificial intelligence systems in generating Chinese information on Ménière’s disease against clinical practice guidelines. Methods: Seventeen questions, adapted from the Key Action Statements of the American Academy of Otolaryngology–Head and Neck Surgery guidelines, were posed to ChatGPT o4-mini-high, Gemini 2.5 Pro, and Grok 3 (51 total responses). Responses were assessed for guideline concordance, communication features, and readability with matched analyses (Cochran’s Q and Friedman tests), with Holm–Bonferroni correction across nine communication characteristics. Results: Correctness rates did not differ significantly among the three models (ChatGPT o4-mini-high: 100%, Gemini 2.5 Pro: 100%, Grok 3: 94.1%; Q = 2.00, p = 0.368). Six of the nine communication characteristics differed significantly, with moderate to large effect sizes (Kendall’s W = 0.26–0.76), including guideline quotation, citation quality, key point emphasis and recommendations beyond the guideline. The proportion of difficult words also differed significantly (p = 0.0033). Gemini 2.5 Pro had a lower proportion of difficult words than ChatGPT o4-mini-high (adjusted p = 0.040) and Grok 3 (adjusted p = 0.0002). Conclusion: High guideline concordance does not necessarily indicate reliable citations, effective communication, or accessible language. These findings reflect responses generated under single-query conditions rather than consistent model performance, highlighting the need for expert oversight in clinical use. Full article

Retrieval-Augmented Generation (RAG) gives language models access to external text and image collections, but it does not make the decisions taken on that content reproducible. Classical Belief–Desire–Intention (BDI) agents provide explicit plans and traceable decisions, although they normally expect beliefs in symbolic form. We connect these capabilities in a hybrid architecture in which RAG updates agent beliefs while BDI plans remain responsible for reasoning and coordination. The plans are implemented in AgentSpeak on JASON/JADE, agents communicate through FIPA-ACL, and probabilistic models are accessed as computational services through the Model Context Protocol (MCP). The orchestrator routes messages but does not derive authorization decisions. Domain agents and the fusion agent produce those decisions through plan execution. We instantiated it for multimodal access control over 167 research posters and compared it with AutoGen and LangGraph baselines on an identical service backend driven by one locally served model, qwen3.5:9b, so that only the control layer differs. Eight campaigns each vary one setting: pipeline length, tool declaration order, the prerequisite and terminal hints, the sampling seed and the fusion policy. They total 4800 baseline runs, with the plan-based arm measured in the two campaigns that have a plan-library counterpart, 300 runs each. Under a favorable configuration, both baselines schedule the workflow correctly and no verdict changes between repeats, yet only 52% and 86% of queries reproduce the same sequence. Their valid-schedule rate falls to 32.0% and 28.0% at nine agents, and presentational edits to the tool declarations move the ordering and, in one arm, the verdict. The plan-based arm returns 100.0% valid schedules and a single sequence in both campaigns. Full article

Low-carbon microgrids are complex energy systems in which photovoltaic (PV) uncertainty, time-varying operating costs, distributed coordination, and information privacy must be addressed simultaneously. This paper couples TCN–BiLSTM probabilistic PV forecasting with privacy-preserving distributed online economic dispatch. Quantile forecasts are converted into a risk-aware net demand by combining the median PV forecast with a lower-side uncertainty reserve. The resulting dispatch problem also includes a step-type carbon-trading cost. To solve the problem when cost gradients are unavailable, we propose a probabilistic PV forecasting-driven differentially private distributed online one-point bandit optimization algorithm (PPF-DP-DOBO). Each generator uses one function-value query per iteration and perturbs its communicated state with Laplace noise. The analysis establishes a per-release differential privacy guarantee, its sequential composition over the dispatch horizon, and an individual dynamic regret bound that explicitly depends on PV forecasting uncertainty. Under bounded weighted path variation and cumulative forecasting uncertainty, the regret is sublinear with order $\mathcal{O}\left(T^{3/4}\right)$. Simulations on a modified IEEE 162-bus system show that the method tracks the risk-aware net demand, preserves the expected privacy–performance trade-off, and yields lower average regret and carbon cost in the evaluated probabilistic-PV setting than in the no-PV case. Full article

Background/Objectives: A cancer diagnosis and its subsequent treatment constitute one of the most psychologically demanding experiences a patient can face, and decades of psycho-oncology research have documented an elevated burden of depression, anxiety, and related conditions in this population. Reported prevalence figures nonetheless vary enormously between studies—a 2011 meta-analysis by Mitchell and colleagues, for instance, arrived at pooled depression estimates roughly half those reported in some subsequent work—complicating clinical decisions about who should be screened and how intensively. The present narrative review with preliminary quantitative synthesis set out to chart this variability directly: to assemble a cross-section of the international literature on depression, anxiety, post-traumatic stress disorder (PTSD), insomnia, and suicidal ideation in oncology patients to quantify how far pooled estimates diverge from one another when treatment setting, tumour type, and assessment instrument are allowed to vary freely, and to lay the empirical groundwork for a subsequent, fully systematic review by the author group. Methods: Consistent with the exploratory aims stated above, eligible primary studies were located through web-based literature searching and citation chaining rather than exhaustive querying of bibliographic databases; the search was therefore not conducted, and is not reported, as a PRISMA 2020-compliant systematic review. Cross-sectional or baseline cohort studies reporting original prevalence figures for at least one of the five outcomes above, in patients with a histologically confirmed malignancy, including adult and paediatric/adolescent patients, who were undergoing or had completed treatment, using a named validated instrument, were retained. Where two or more studies addressed the same outcome, a random effects pooled proportion (DerSimonian–Laird estimator, Freeman–Tukey double arcsine variance stabilisation) was computed in Python as an illustrative, not definitive, summary statistic, alongside Cochran’s Q and I². Results: Twelve studies conducted across nine countries were retained, contributing fourteen outcome-level data points. The illustrative pooled prevalence of depression was 45.6% (95% CI: 29.5–62.3%; k = 3; I² = 91.2%) and of anxiety 44.1% (95% CI: 32.7–55.8%; k = 4; I² = 90.5%)—both considerably above the double-digit figures reported in earlier interview-based meta-analyses, reflecting the self-report screening instruments and, in several instances, lower-income treatment settings represented in the present sample. Pooled PTSD prevalence was 17.0% (95% CI: 7.3–29.7%; k = 3; I² = 91.8%) and pooled insomnia prevalence 38.0% (95% CI: 24.5–52.6%; k = 3; I² = 86.7%). A single Chinese multicentre study of 509 women with advanced breast cancer reported suicidal ideation in 22.8% of respondents; this figure is reported descriptively only, is not pooled, and should not be extrapolated to oncology patients generally. Conclusions: Even acknowledging its exploratory character, this synthesis reinforces a conclusion already well established in psycho-oncology: a substantial minority, and on some metrics close to half, of patients receiving cancer treatment report clinically significant psychological symptoms, with figures highest for depression and anxiety in the self-report studies assembled here. The magnitude of between-study heterogeneity observed for every outcome, however, argues strongly against treating any single percentage—ours included—as a stable population estimate; tumour type, treatment phase, country income level, and instrument choice all plausibly move the figure by a considerable margin, and disentangling these influences is the explicit purpose of the fully systematic review the author group is now undertaking; the present figures should be read as hypothesis-generating inputs to that forthcoming systematic review, not as stand-alone prevalence estimates suitable for direct clinical or policy use. Full article

Frequent extreme floods and the limitations of traditional manual consultation—information latency, knowledge fragmentation, and experience-dependent reasoning—create an urgent need for intelligent flood control decision support. To address the complex hydrological conditions and urgent consultation demands involved in flood control decision-making at the Xiaolangdi Reservoir, this study develops an intelligent decision support system that enhances emergency response capabilities through the integration of advanced artificial intelligence technologies. The system integrates a multi-agent architecture with Retrieval-Augmented Generation (RAG) and Large Language Models (LLMs) to create a comprehensive consultation platform. It incorporates multiple specialized agents, including data analysis agents, consultation reasoning agents, and consultation querying agents, which collaboratively process real-time hydrological data, historical flood records, and operational constraints through carefully designed prompt chains. The RAG component enables efficient retrieval of relevant historical cases and technical specifications from the knowledge base, while the LLM engine generates contextualized consultation suggestions and operational recommendations. Practical deployment demonstrates that the system significantly improves the timeliness of flood control decision-making by substantially reducing consultation time. The prompt engineering framework, incorporating domain-specific templates and adaptive reasoning mechanisms, ensures that the generated consultation advice complies with reservoir operational standards and safety protocols. By providing a scalable intelligent consultation platform, this research substantially enhances decision-making efficiency and reliability during emergency flood events, promoting digital transformation in flood management practices. Full article

To address the attenuation of shallow fine-grained structural information during deep feature propagation and the representational conflict between classification and regression in remote sensing tiny object detection, this paper proposes a dual asymmetric detection framework, termed $A^2$-Det. Built upon YOLO11n, the detection pyramid is shifted toward the high-resolution P2–P4 levels by introducing a P2 detection branch and removing the original P5 stage, thereby reducing the loss of fine-grained spatial details caused by successive downsampling. On this high-resolution feature basis, a Query–Key–Value (QKV)-guided Asymmetric Spatial Feature Enhancement module (Q-ASFE) is deployed at the P2 and P3 stages, where direction-sensitive asymmetric convolutions are combined with lightweight QKV-based contextual modulation to strengthen tiny object structural responses while suppressing complex background interference. Furthermore, an Asymmetric Coordinate–Semantic Decoupled Head (ACS-Head) performs differentiated modeling for the semantic selection required by classification and the coordinate-sensitive representation required by regression, thereby alleviating task-specific representational conflict. Through the progressive coordination of high-resolution feature preservation, shallow-feature enhancement, and task-specific prediction, $A^2$-Det improves mAP₅₀ by 5.8 percentage points over YOLO11n on the VisDrone dataset. Consistent improvements on the USOD and RSOD datasets further demonstrate its effectiveness and cross-scene generalization. Full article

Background/Objectives: Artificial intelligence (AI) tools are used in clinical education and patient communication. Concordance with American Academy of Otolaryngology–Head and Neck Surgery Foundation clinical practice guidelines (CPG) and overall readability remain poorly defined. Methods: Over 2 months, ChatGPT, Google Gemini, and Google Search AI were queried using a single LF prompt and a layered set of sub-questions. Responses were scored with a 12-point rubric from the CPG plain-language tonsillectomy guideline. Readability was assessed using Flesch Reading Ease (FRE) and Flesch-Kincaid Grade Level (FKGL). Temporal trends and between-group differences were analyzed using regression and repeated-measures methods. Results: Layered prompts demonstrated greater alignment with guideline subtopics than LF (89.5% vs. 71.4%, p < 0.001). For LF queries, mean concordance was 74.5% (95% CI 72.0–77.0) for Google AI, 71.3% (68.9–73.7) for Gemini, and 68.4% (66.0–70.9) for ChatGPT. Overall, Google AI outperformed Gemini and ChatGPT on pairwise comparisons (p = 0.001). Google AI demonstrated improvement in concordance (slope = 0.03, p = 0.003), while Gemini showed no change (p > 0.05) and ChatGPT worsened over time (slope = −0.03, p < 0.001). Readability differed by strategy and model. Layered prompts produced more accessible text (FKGL 44.6, FRE 10.3) compared with LF outputs (FKGL 35.2, FRE 11.3) and the CPG (FKGL 10.6, FRE 41.7). IRR was high (ICC = 0.96). Conclusions: Model type and prompt structure shaped the quality of AI-generated tonsillectomy guidance. Google AI consistently outperformed ChatGPT, with Gemini intermediate, and uniquely showed temporal improvement. Layered prompting produced more CPG-aligned and readable responses across all models. Readability was consistently above a 7th grade level. Full article

3D Gaussian Splatting (3DGS) enables real-time photorealistic scene reconstruction, yet its segmentation tasks suffer from two critical flaws: poor 3D consistency (e.g., blurred instance boundaries and unstable cross-view semantic association) and insufficient structural awareness near ambiguous object boundaries. To address these issues, this paper proposes a semantic-guided 3D Gaussian segmentation framework. We first process multi-view RGB images and semantic masks to generate 2D semantic codes and initial Gaussian parameters. We then build a Gaussian-level joint representation by combining flow-aligned appearance cues, CLIP-derived semantic descriptors, and explicit Gaussian geometry. Semantic-guided adaptive decomposition identifies boundary-sensitive Gaussians through cross-view boundary statistics and semantic uncertainty, while multi-view voting further refines instance boundaries. Finally, 3D global optimization unifies instance association and labeling. Experiments on ScanNet and SPIn-NeRF show improved results under the reported evaluation protocol and the stated subset definitions. In particular, on the ScanNet benchmark, compared with the evaluated 3DGS baseline InstanceGaussian, SG-AGS improves mAcc@0.25 by 16.2 points for category-agnostic 3D instance segmentation and by 25.6 points for text-query-based open-vocabulary labeling of segmented 3D instances. These results support the usefulness of SG-AGS on the two evaluated benchmarks and suggest potential value for object-level scene querying and digital-twin inspection, while broader generalization to external captures and other scene domains still requires further validation. Full article

Macedonian is a low-resource language for automatic speech recognition: annotated speech data are scarce, dialectal variation is substantial, and existing evaluations focus almost entirely on read speech in Standard Macedonian. This paper presents a unified pipeline that converts Macedonian speech, including regional dialects, into accurate transcripts and structured, queryable knowledge, a capability required for applications such as searchable parliamentary archives, broadcast transcription and subtitling, and dialectological documentation. Parameter-efficient adaptation of Whisper large-v3-turbo via low-rank adaptation is evaluated against strong zero-shot and language-specific baselines on four newly curated dialect corpora (Ohrid, Veles, Tikvesh, and Gostivar) and three Standard Macedonian corpora, two of which were collected for this work. The adapted model reduces word error rate by 57–70% relative to the strongest zero-shot baseline and by 33–66% relative to the language-specific BUKI Whisper 2.0 model on dialectal speech, with comparable improvements over zero-shot baselines on standard Macedonian speech, while updating only about 0.7% of parameters. Beyond transcription, the pipeline adds speaker diarization with cross-recording speaker linking and a graph-based retrieval-augmented generation component that enables speaker-, topic-, and time-aware querying of diarized transcripts, evaluated on long-form Macedonian parliamentary recordings. Together, these results establish parameter-efficient adaptation, speaker-aware processing, and graph-based retrieval as a practical and transferable framework for transforming under-resourced speech into accessible, structured knowledge. Full article

Aerial scenes exhibit a mismatch between where computation is spent and where information resides: targets occupy a small, unevenly distributed fraction of the frame, yet detectors process every region at identical resolution. Tiling and clustering pipelines exploit this sparsity, but the rule deciding which regions deserve magnification is hand-designed and never observes the detection loss. AFWD-Net removes that separation. A coarse density estimator, run on a heavily downsampled copy of the frame at negligible cost, produces a spatial prior over likely target locations; Gumbel top-K relaxation and a spatial-transformer sampler convert this prior into a small set of high-resolution crops whose centers and scales are differentiable, so the gradient that measures detection quality also determines where resolution is allocated. Within each crop, a learnable wavelet decomposition separates approximation from directional detail sub-bands and enhances the latter under edge supervision confined to annotated regions, restoring the high-frequency structure that stride downsampling erases, precisely the cue separating a genuine sub-32-pixel target from background clutter. Detection queries are seeded from the same density field rather than a fixed lattice, and crop-level predictions are merged by density-calibrated soft suppression. On VisDrone2019-DET the model attains 49.3% mAP@0.5, 30.4% mAP@[0.5:0.95] and 22.8% APS at 15.7 M parameters; a distilled 7.2 M student remains competitive with all compared methods. Experiments on UAVDT, AI-TOD-v2 and SODA-A assess cross-dataset generalization. Full article

AI assistants are increasingly integrated into digital financial services; however, relying on generative language models for response generation can introduce challenges related to response reliability, data exposure, and resource requirements. This study presents a hybrid data-grounded architecture that selectively uses generative AI while maintaining reliable natural language interaction. The proposed task-oriented system combines semantic intent matching, slot and entity extraction, reconciliation, and response routing. Instead of using the generative model to describe every response, the architecture constructs template-based responses directly from structured institutional data whenever sufficient information is available and invokes the generative language model only as a fallback. The architecture was evaluated using a fixed set of 100 realistic user queries, including diverse expressions and spelling variations, across 14 iterative system refinements while the underlying AI models remained unchanged. Response correctness was manually assessed against the underlying data and classified as correct, unanswered, or incorrect. Correct responses increased from 51% to 97%, whereas incorrect responses decreased from 33% to 0%. Simultaneously, language-model-generated responses decreased from 34% to 10%, whereas data-grounded deterministic responses increased from 66% to 90%. These findings indicate that system-level architectural refinement can substantially improve response reliability while reducing dependence on generative response production. However, the observed reduction in generative-model dependency should be interpreted as a potential resource-conscious architectural strategy rather than evidence of a measured environmental benefit. Full article

Background/Objectives: Large language models (LLMs) are increasingly explored for clinical decision support and digital health applications. However, reliable diagnostic assistance remains challenging for low-resource medical languages such as Turkish due to limited source grounding, transparency, clinical safety, and localized medical knowledge. This study presents TurkishMedLLM, a source-grounded and safety-aware Turkish medical LLM designed for clinician-supervised diagnostic decision support, symptom interpretation, and patient care. Methods: The methodology integrates multi-source Turkish medical data ingestion, schema standardization, duplicate removal, quality filtering, supervised fine-tuning, embedding generation, vector indexing, Qwen3-8B fine-tuning, retrieval-augmented generation (RAG), and multi-layer evaluation. The system was evaluated using retrieval metrics, ROUGE, RAGAS, DeepEval, and clinical safety assessments. As a use case, TurkishMedLLM was integrated into the AI-based Diabetes Care (AIDCare) mHealth platform, which supports patient queries related to lifestyle management, symptoms, diagnosis, and treatment of diabetes. The system generates safety-aware responses with clinician-in-the-loop validation before delivery through the mobile application. Results: After pre-processing, the final dataset comprised 232,926 unique documents, including 210,791 Turkish medical question–answer pairs and 22,135 hospital medical articles. The retrieval module achieved Hit Rate@1, Hit Rate@3, and Hit Rate@5 of 94.67%, 98.67%, and 100.00%, respectively, indicating consistent retrieval of clinically relevant evidence. QLoRA fine-tuning achieved a validation loss of 0.9373 and ROUGE-1, ROUGE-2, ROUGE-L, and ROUGE-Lsum scores of 0.8338, 0.5476, 0.7836, and 0.7372, respectively. The fine-tuned RAG system achieved RAGAS faithfulness and response relevancy scores of 0.91 and 0.88, respectively, while DeepEval achieved an answer relevancy score of 0.90. Clinical safety assessment achieved a caution score of 0.93, indicating generally evidence-grounded and clinically cautious responses for symptom interpretation and diagnosis-related patient support. Conclusions: Combining retrieval grounding, parameter-efficient fine-tuning, and multi-layer safety evaluation provides a promising approach for clinician-supervised medical AI in Turkish. TurkishMedLLM demonstrates potential for symptom interpretation, differential diagnostic support, and trustworthy digital health applications. Full article

Reliable positioning underpins connected and autonomous vehicles, delivery robots, public-transport monitoring and infrastructure inspection in smart cities. In urban canyons, covered roads and tunnels, satellite navigation can degrade, making camera-based visual place recognition (VPR) a fallback. Most VPR benchmarks, however, evaluate whether a correct place appears in a ranked list, whereas an operational positioning component must issue a coordinate-or-abstain with stated availability. We present an estimator-aware audit framework that tests whether retrieval results support such positioning-service claims. It separates reference-map opportunity, candidate retrieval, coordinate selection, acceptance and temporal aggregation, matching each comparison to its query population and estimator type. We apply the framework to leakage-controlled, route-conditioned KITTI driving sequences and prospectively test the frozen pipeline on Oxford RobotCar, NCLT and St Lucia traversals. Retrieval performance alone did not establish reliable coordinate output: no tested causal acceptance policy demonstrated the prespecified combination of high output coverage and low positioning risk, and the external service gate failed on all three independent datasets. Stage-wise analysis distinguished missing map opportunity from candidate-generation, candidate-selection and temporal-aggregation failures. For smart-city development, the findings show that resilient mobility infrastructure requires joint evaluation of output availability, positioning risk, map design and cross-domain robustness before deployment. Full article

Background: Plant-derived natural products are a productive antiviral scaffold source, yet secondary metabolite libraries remain underexplored against HIV-1 amid extensive target-structure redundancy. This study presents a cluster-guided framework for HIV-1 inhibitor prioritization from the KNApSAcK database. Methods: A total of 64,166 SMs were converted to SMILES and queried against BindingDB to identify reported HIV-1 integrase, protease, and reverse-transcriptase associations. A total of 295 HIV-1 protein sequences were aligned and partitioned using DPClusSBO; representative structures (6VDK, 1MUI, 6ELI) per cluster were docked against cluster-mapped SMs using SMINA. Prioritized SMs were evaluated with SwissADME and benchmarked against ChEMBL HIV-1 inhibitors. Results: Clustering resolved three non-overlapping groups (integrase, n = 135; protease, n = 117; reverse transcriptase, n = 42), mapping 285, 124, and 410 SMs, respectively. Predicted docking scores ranged from $-18.91$ to $-4.72$ kcal/mol (integrase), $-26.54$ to $-7.87$ kcal/mol (protease), and $-26.83$ to $-4.59$ kcal/mol (reverse transcriptase). ADME-prioritized reverse-transcriptase compounds scored more favorably than the matched NNRTI reference set (p < 0.001), requiring experimental confirmation of binding affinity. DUD-E enrichment validation showed strong discriminative validity for integrase and protease (ROC-AUC 0.85, 0.78) but not reverse transcriptase (ROC-AUC 0.53), consistent with weaker pose reproduction for the latter. Conclusions: The framework reduced target redundancy and computationally prioritized natural product candidates for HIV-1 as hypothesis-generating predictions requiring experimental validation. Full article