Many industrial processes rely on climate control systems whose thermal or hygrothermal inertia allows for a degree of temporal flexibility in electricity consumption, which can be exploited through load-shifting strategies toward lower-price time slots. This work presents an Internet of Things (IoT) system for electricity cost optimization in this type of process. The system integrates a data acquisition layer for industrial process and electricity market data, a data-driven process model automatically extracted from historical data, and a simulation engine with coupled humidity–power flow. Based on this engine, three optimization strategies (E1–E3) of increasing complexity are defined, designed with a viewto deployment on embedded devices. The system is validated on an industrial drying chamber in the cheese sector over one full annual cycle, comprising 355 evaluated days between February 2025 and February 2026. The best strategy achieves annual savings of 44.65%under the most aggressive configuration evaluated (three saving windows of four hours per day), and between 11.3% and 36.0% under more conservative ones. The evaluation is complemented by a leave-one-month-out cross-validation and by an external validation on a second drying chamber with different setpoints and load. The results show that simulation-based strategies substantially outperform purely price-based selection, and that the strategy with prior pruning (E2) constitutes an effective approximation of the exhaustive search (E3), with a difference of at most 0.25 percentage points and a computational complexity compatible with embedded hardware.
This paper presents a scalable FPGA-based architecture for the Finite-Difference Time-Domain (FDTD) method in computational electromagnetics. The proposed design addresses key limitations of existing FPGA accelerators, including on-chip memory constraints, external memory bandwidth bottle-necks, and limited multi-node scalability. The architecture combines synthesis-time tiling, fully pipelined dataflow execution, burst-based memory transfers, and optimized fixed-point arithmetic. The tiling strategy enables exact partitioning of the computational domain while fitting per-tile buffers within BRAM and URAM resources. The design is further extended to a multi-node cluster, where FPGA devices process independent sub-domains and exchange boundary data synchronously at each time step, enabling scalable parallel execution. Experimental results on an AMD Kria KV260 platform quantify the impact of each optimization in terms of performance, resource utilization, and efficiency, highlighting the trade-offs between memory usage and parallelism in FPGA-based FDTD accelerators.
The color quality control of general-purpose polystyrene productions is assessed by visual inspection currently in industry. In order to overcome the limitations of this method and find a more scientific and robust way to classify the color grades of plastic samples, a new approach is explored. We first discuss several unsupervised machine learning clustering algorithms and analyze their advantages and drawbacks. The major color measurement methods are also evaluated. A device is proposed to measure the color temperature and illuminance of plastic samples. These measured color characteristics have the ability to recognize the subtle difference in color, overcoming the limitations of existing color measurement instruments. Based on these measured quantities, six unsupervised clustering algorithms are applied to verify the performance of clustering including K-Means, Spectral clustering, Agglomerative clustering, BIRCH, Bisecting K-Means, and Gaussian mixtures model. Five different metrics are used to evaluate clustering results. The obtained clustering regions basically match well with the reference labels. The results demonstrate that unsupervised clustering can effectively support color quality control and grading in plastic manufacturing.
Accurate environmental perception is critical for autonomous vision systems, requiring joint estimation of object locations, scene geometry, and object orientation in real time. Existing monocular approaches address detection, depth estimation, and orientation as isolated tasks, leading to high memory overhead and spatial inconsistency that make them unsuitable for real-time use. No lightweight single-pass RGB pipeline currently integrates all three tasks at real-time frame rates on consumer hardware.This work proposes a integrated monocular framework that simultaneously performs object detection, dense depth estimation, and orientation regression from a single RGB image. The architecture integrates a computationally efficient detection module with transformer-based depth estimation and a regression-based orientation module, enabling multimodal scene understanding without relying on LiDAR or stereo inputs. Shared visual feature representations across detection and regression tasks enhance spatial coherence and cross-task consistency. Extensive experiments on the BDD100K and KITTI datasets demonstrate robust performance, achieving a mean Average Precision (mAP) of 45.3% for detection, a depth Mean Absolute Error (MAE) of 1.85 m, and orientation MAEs of 4.7° (yaw) and 3.2° (pitch). The complete system operates in real time at 25–30 frames per second, highlighting its suitability for resource-constrained robotic and autonomous driving applications.
This paper presents a broadband reconfigurable low-noise amplifier (LNA). With the proposed reconfigurable topologies, the design alleviates the constraints between bandwidth and Noise Figure (NF). The proposed four topologies include: 1) A low-noise oriented input stage that employs four bandwidth extension techniques, i.e., high-resistance shunt-shunt feedback, L-C-R load peaking technique, inductive-peaking technique, and zero-point tuning technique to achieve wideband impedance matching and NF optimization; 2) Switchable diplexer that splits signals into two frequency bands. It is frequency reconfigurable to provide band overlap and relieve pressure on gain smoothing in the post-stages; 3) Gain smoothing stages that employ inductive shunt feedback to manipulate a pair of conjugate poles to compensate for the gain variation; 4) A high-isolation SPDT switch to combine the signals and avoid possible oscillations. The proposed reconfigurable LNA is fabricated in a commercial 0.15-μm GaAs pHEMT process. Experimental results show that the LNA achieves a wide BW from 2.5 to 18 GHz with a favorable NF of 1.38–1.97 dB. The LNA can be reconfigured into low-band (2.5–8 GHz) and high-band (6–18 GHz) modes. At low-band mode, the gain is 20.3–22.1 dB, the NF is 1.38–1.53 dB, and the output P1dB points are 2–7.4 dBm. At high-band mode, the gain is 20.2–22.7 dB, the NF is 1.64–1.97 dB, and the output P1dB points are 3.5–7.2 dBm. The input and output impedance are matched well.
Wireless power transfer (WPT) has improved the development of active implantable medical devices (AIMDs) by enabling them to operate wirelessly, allowing uninterrupted treatment to patients without direct intervention. Inductive power transfer (IPT) is the most established technology, composed of magnetically coupled coils to form inductive links. However, delivering efficient, stable, and safe power to an implant is not trivial. This paper presents a methodological review on the design of IPT systems applied to AIMDs. The fundamental concepts of IPT are presented, including inductive link modeling, compensation networks, winding methods and the influence of frequency. Design and optimization techniques at the system level are discussed, detailing the most common topologies and incorporating electromagnetic (EM) safety and biocompatibility aspects. The main control methods for IPT systems to achieve stable output power are detailed and categorized based on their implementation in the primary, secondary, or through external communication. The advantages and disadvantages of each method are discussed. Finally, the literature findings are presented for various applications, such as pacemakers, neurostimulators, cochlear, or retinal implants. Both commercial industry solutions and research works are detailed. These works are compared quantitatively, including a metric to evaluate the quality of the proposed IPT systems.
Large language models (LLMs) increasingly generate executable optimization code, yet evaluations often rank programs by objective value, overlooking deployment-relevant properties such as validity under structural constraint changes, failure localization, repairability, and component compatibility. We introduce GenSE-Scheduler, a deployment-oriented framework for evaluating LLM-generated parallel-machine scheduling code. It compares monolithic schedulers with four-stage modular pipelines under typed contracts, sandboxed execution, a failure taxonomy, stage-swap repair, and pre-composition compatibility prediction. We evaluate 1,350 artifacts from four locally served open-weight model families across 1,039,600 artifact–instance pairs. Modular pipelines exhibit substantially greater outcome-level behavioral diversity than monolithic artifacts (Cliff’s δ = +0.799 for cluster entropy; δ = +1.000 for pairwise regret distance). However, modular generation does not meet the pre-registered regret-equivalence criterion (Δ = +0.565 > ε = 0.05), although the effect is negligible (δ = +0.044) among passing in-distribution pairs. Among failing modular pipelines, 97 of 98 (99.0%) admit at least one valid single-stage substitution, but this repairability is contract-level: the resulting swaps rarely meet tight scheduling-quality tolerances. Repairability is also strongly stage-dependent, with the assign stage as the main bottleneck. A static Module Compatibility Predictor reaches macro-F1 = 0.872 for predicting compatible swaps before composition. Machine-eligibility stress tests—using a constraint that is present in the generation prompts but never instantiated in the C1–C4 calibration regime—yield a 0% pass rate in both an isolated eligibility class and a combined release-date/setup-time/eligibility hold-out, which we interpret as a joint calibration-coverage and serialization-contract boundary, not evidence that LLMs cannot implement eligibility constraints when those constraints are explicitly instantiated and exercised. Against a classical greedy min-completion-time list-scheduling baseline, neither architecture is makespan-competitive: the heuristic attains far lower regret and is matched or beaten by well under 1% of passing artifact–instance pairs. Overall, modular generation does not dominate monolithic generation in raw scheduling quality; rather, modular decomposition makes deployment-relevant properties measurable.
Diffusion-model watermark detectors are usually evaluated under fixed attack distributions, although deployed systems face continual attack shift. When out-of-distribution attacks reduce detection quality, full retraining is too slowfor latency-sensitive verification services. This paper formulateswatermark detection recovery as a task-level adaptation problem and proposes a meta-watermarking framework that learns a shared detector initialization via first-order Model-Agnostic Meta-Learning (FOMAML) on 10-dimensional ROBIN ring-frequency features. Recovery is measured by the minimum adaptation step k∗ needed to satisfy a fixed TPR/FPR operating point. Across six data scales (16–512 images), seven attack types, and four methods, the meta-initialized 11-parameter logistic detector achieves zero-shot recovery (k∗ = 0) on unseen rotation attacks at N = 128, with TPR=0.875 and FPR=0.063, whereas full retraining requires k∗ = 2. Threshold-only recalibration fails at this scale on rotation, indicating that learned initialization remains beneficial at moderate data budgets. After feature extraction, the detector classifies in 0.079 ms per image, about 43,000× smaller than the shared DDIM inversion stage that dominates end-to-end latency. Exploratory Spearman trends over the five single-seed scales (N = 16–256) suggested faster recovery with scale, but including the N = 512 stress test (n = 6) removes this correlation, and the multi-seed recovery rates are non-monotonic, so the trend does not generalize across all tested scales. The results are therefore consistent with meta-initialization as a practical recovery mechanism for the tested rotation attack shift, while recovery under high-noise conditions remains limited.
This paper proposes a HILS-based data-driven curtailment weighting control framework for wind farm operation under grid-imposed output constraints. The main objective of the proposed framework is to extract wind farm operating data from a real-time hardware-in-the-loop simulation environment and utilize them for data-driven turbine-level power prediction and objective-function-based curtailment optimization. In the proposed framework, an RTDS-based wind farm model, wind farm management system, and wind turbine controllers are interconnected through Modbus TCP/IP communication to generate and collect operating data under curtailment conditions. The extracted data are processed and used to train a CNN-BiLSTM model for turbine-level power prediction, which provides available power information for the curtailment weighting control module. Unlike conventional forecasting-oriented approaches, the proposed method links data-driven prediction with control-oriented decision-making by incorporating an operational objective into the curtailment allocation process. As a representative case, internal power loss reduction is adopted as the objective function for determining turbine-level curtailment weights. Simulation results show that the proposed framework generates feasible active power references, reduces curtailment-related losses compared with conventional methods, and maintains turbine availability under constrained operating conditions.
Industrial IoT predictive maintenance demands real-time anomaly detection under tight resource and interpretability constraints, while monolithic LLM-based systems remain impractical for on-site deployment. We introduce HAMA (Hierarchical Adaptive Multi-Agent Architecture), in which ‘‘adaptive’’ refers strictly to online adjustment of policy parameters, not of agent structure. Edge agents perform lightweight statistical pre-filtering; K=3 Fog nodes run a five-model detection ensemble with consensus voting and federated-style parameter aggregation; Cloud agents adapt consensus weights and thresholds online via Proximal Policy Optimization (PPO) and audit alerts with SHAP attributions, which ground a locally hosted small language model (Llama-3.2-1B) that generates operator-facing explanations. We evaluate HAMA against static and rule-based adaptive baselines on a corrected Boiler Emulator benchmark, a full-length chronologically splitWind Turbine SCADA series with predictive 60-minute-ahead labels, and NASA C-MAPSS with genuine run-to-failure RUL labels (LSTM regressor: MAE 11.2 / RMSE 16.0 cycles, comparable to published results). Across five seeds, PPO-based adaptation is statistically indistinguishable from both baselines on detection F1 (p > 0.5 in every pairwise test) - a null result we report directly - while end-to-end detection latency (0.8–2.0 ms, measured on commodity CPU hardware) stays well within the 100 ms real-time budget, and robustness under noise, missing data, and reduced prevalence matches the static baseline. HAMA thus demonstrates that tiered deployment, real online adaptation, multi-node aggregation, and locally hosted explainability can be integrated into one measured, reproducible system without sacrificing detection quality - an architectural contribution whose statistics and resource footprints are all regenerated from released artifacts. Implementation: https://github.com/HySonLab/AgentIoT.
Smart contracts increasingly support high-value and governance-critical blockchain applications, making precise program understanding important for reliable analysis and tooling. Many existing analysis tools rely on task-specific pipelines in which extracted program knowledge is not readily available as reusable and independently queryable semantic data. This paper presents SolOnto, an executable ontology that provides a compiler-grounded semantic representation of Solidity contract structure and selected execution-related semantics. Compared with the ontology and knowledge-graph approaches reviewed in this study, SolOnto distinguishes itself by automatically materializing compiler-produced abstract syntax trees (ASTs) as ontology-aligned Resource Description Framework (RDF) instances and supporting SPARQL-based semantic retrieval. The graphs represent core entities and selected semantics, including statement ordering, state-variable access, and external interactions. Structural semantics are assessed qualitatively, whereas execution-related semantics are evaluated through an implementation-conformance assessment using precision and recall over competency-question-driven queries within the selected contract set. The pipeline is further evaluated using publicly verified contracts from Ethereum Mainnet, Arbitrum, and Optimism Mainnet, including flattened Solidity sources and multi-file Standard JSON artifacts. An evaluated use case examines ontology-supported natural-language-to-SPARQL interaction through constrained query generation, basic query-structure validation, and predefined fallback mechanisms. The evaluated interface is characterized as a constrained hybrid pipeline whose operationally successful responses depend predominantly on predefined fallback templates. Accordingly, the reported end-to-end operational success rate reflects the complete pipeline rather than the standalone translation capability of the LLM. Within the evaluated scope, the findings support the feasibility of SolOnto as a reusable semantic foundation for transparent and queryable smart-contract inspection.
Precision agriculture relies on accurate computer vision frameworks to support remote crop monitoring and field management. However, training accurate semantic segmentation models is heavily constrained by the scarcity of annotated agricultural imagery. This study examines annotation segmentation methods for productivity and accuracy under sparse data by comparing two well-known methods: a classical machine learning approach (Random Forest) and a deep learning architecture (DeepLabv3+). The models are evaluated on multiple agricultural datasets representing different sensing modalities and spatial scales. Each of these includes UAV imagery and ground-level field images. Results have shown that DeepLabv3+ produces more spatially clear segmentation and improved crop–weed delineation in complex scenes, while Random Forest remains a computationally faster and efficient algorithm at the pixel level in simpler annotations. The results demonstrate that combining sparse manual labeling with model-assisted prediction can reduce annotation effort while still providing reliable field information for farmers and precision agriculture applications.
Organizations increasingly rely on heterogeneous digital performance data, including KPIs, activity logs, competency ratings, and 360-degree feedback, to support employee appraisal and development. However, transforming such evidence into interpretable, personalized, and governable narrative feedback remains an open challenge. Natural Language Generation (NLG) and large language models (LLMs) offer a promising basis for this transformation, yet no prior review has systematically examined their applicability to automated and personalized employee evaluation systems. This study presents a systematic literature review conducted using PRISMA-informed procedures. Searches across IEEE Xplore, ACM Digital Library, ScienceDirect, Scopus, and Web of Science identified studies published between January 2015 and September 2025. After screening and eligibility assessment, 98 papers were retained and analyzed across five dimensions: task families, methodologies, data sources, evaluation practices, and deployment challenges. As employee evaluation is an emerging application area, the retained studies are drawn primarily from adjacent domains, including education, healthcare, and coaching, and the review therefore synthesizes transferable components rather than direct evidence from workplace settings. The review identifies four task families relevant to employee evaluation: automated feedback generation, evaluative judgment, report and template generation, and personalized developmental guidance. Prompt-based LLM approaches dominate the methodological landscape. Personalization remains relatively shallow and is rarely validated against downstream outcomes. Governance concerns, particularly those related to privacy, fairness, and deployment-time explainability, are inconsistently addressed, while the evaluation methodology remains the main bottleneck in the field. In general, the necessary building blocks exist across adjacent domains but remain fragmented, and their transfer to workplace evaluation will require domain-specific adaptation. Progress will require integrated architectures combining evidence grounding, role-sensitive personalization, standardized evaluation protocols, and institutional accountability.
Frota 360 is a robot fleet management (RFM) framework designed to support autonomous inspection in oil and gas (O&G) facilities, where operators must coordinate heterogeneous robots under strict safety, connectivity, and cybersecurity constraints. Existing commercial solutions are typically cloud-centric, vendor-specific, or weakly integrated with mission planning and inspection workflows, while opensource frameworks such as Open robotics middleware framework (Open-RMF) primarily focus on traffic negotiation and basic task orchestration. This paper presents an on-premise, physically isolated architecture that extends Open-RMF with an O&G-oriented interoperability and supervision layer, enabling coordinated multi-robot inspection. As a work in progress, the proposed system introduces a vendor-agnostic abstraction for legacy and modern robots. We present an early iteration that extends RMF-Web with a supervisory dashboard tailored to the operational requirements of Petrobras, the leading O&G company in Brazil. The envisioned interface incorporates role-based access control via Keycloak, live streaming of camera and perception data, and 2D/3D visualization of robot states aligned with facility maps and computeraided design (CAD) models, supporting situational awareness and operator-in-the-loop supervision in industrial environments. The ongoing development of Frota 360 is motivated by an initial in-situ analysis at Petrobras facilities using ANYmal, Spot, and the Taurob Inspector under commercial fleet management systems to identify architectural and operational limitations. We validate the architecture through a phased methodology whose results follow a middleware comparison of Zenoh against data distribution service (DDS) implementations over the experimentalWi-Fi mesh; quantitative gains from large-scale CAD loading optimizations; laboratory monitoring-task trials with two real wheeled robots, including sim-to-real timing, that exercise dispatch, live supervision, and photo generation under an isolated stack; and physics-based Gazebo experiments of multi-robot inspection timing in an offshore digital twin. Together, these experiments de-risk the communication, environment-modeling, and coordination choices that advance Frota 360 toward a deployable on-premises RFM for O&G inspection.
Autonomous driving systems dealing with complicated traffic environments need to not only detect objects in the traffic, but also identify and prioritize dangerous road users. Traditionally, object detection models have focused on localization and classification performance, while treating all detected objects as equally important. But, in real-world driving situations, understanding object risk in context is needed for safer decision-making. This paper proposes a Risk-Aware YOLOv8 model that combines object detection and risk assessment into a single architecture for autonomous driving applications. The proposed approach extends the YOLOv8 detection system by including another branch of risk prediction that estimates the relative risk of each object detected. Risk scores are estimated using a multi-factor model taking safety severity, distance risk, driving corridor relevance, and failure-aware confidence estimation. This approach combines object detection with risk-aware prioritization of road entities. Experiments were conducted with the DriveIndia dataset [4] containing various Indian road scenarios and 25 objects. The model produced precision of 0.768, recall of 0.570, mAP@50 of 0.660, and mAP@50–95 of 0.569, indicating good detection performance for a range of traffic participants and road infrastructure elements. The results suggest that the estimation of the risk at various stages of the detection process could enhance situational awareness and provide additional decision support for autonomous driving. The proposed Risk-Aware YOLOv8 approach looks promising as a direction for safety-focused perception that can detect objects with accuracy and prioritize objects in real-world situations.
Micro Electro Mechanical Systems (MEMS) piezoresistive pressure sensors, widely used in industry, aerospace and automotive applications. However, it will exhibit low accuracy and less stability due to high pressure and temperature fluctuations. In order to overcome the aforesaid problems by developing a unique piezoresistive pressure sensor design is introduced in this work. The SiC diaphragm consisting of a SiC circular rib used to strengthen the stiffness of diaphragm, stable the deflection and improve temperature stability. The primary function of rib is to lessen excessive center deflection and more evenly distribute stress throughout the diaphragm upon the application of pressure. The pressure range of 0-40 MPa was applied to analyze the sensor performance with and without a circular rib. A mathematical model was developed by using theory of thin plate and small deflection. This model was simulated inMATLAB to examine the sensor deflection, stress, output voltage and sensitivity. COMSOL Multiphysics Finite Element Analysis (FEA) was utilized to verify the mathematical findings and Machine Learning (ML) methods used for prediction. The performance of the sensor can be examined in under high pressure. With high sensitivity (4.41 mV/V/MPa) and low nonlinearity (2.23%), the Artificial Neural Network (ANN) model predicts sensor behavior with accuracy (98%). Compared to previous research, an ideal rib coverage of 3 − 5% is found, which enhances performance and permits effective smart sensor design.
This paper presents a method to correct errors in real-time voltage measurements in a power distribution network based on advanced metering infrastructure (AMI) data and to determine the real-time voltage condition of all buses from the corrected voltage. It focuses on two aspects. First, for voltage measurement in medium-voltage distribution networks, the causes and characteristics of errors in the commonly used capacitor voltage divider method are analyzed, and such errors are shown to exhibit identifiable patterns. On this basis, coefficients are derived offline to correct errors in voltage measurements using seasonal and time-of-day patterns obtained from AMI data. Second, this study addresses issues arising from modeling distributed loads in actual distribution systems as lumped-load representations and from error accumulation in voltage estimation due to reliance on pseudo-measurements. Accordingly, a voltage-estimation method is proposed for online operating conditions, in which a virtual sectional voltage-drop reference point is shifted according to the corrected voltage measurements. The proposed method was implemented in the advanced distribution management system of the Korea Electric Power Corporation, and a case study was conducted by using long-term data from an actual distribution network. The method was highly effective in reducing both average voltage-estimation error and instantaneous errors. In particular, the error reduction was significant during periods when the distributed generation output had a strong influence. The proposed method is directly applicable to actual distribution networks because it fully accounts for the non-real-time constraints of AMI data while excluding additional equipment and assumptions currently introduced into distribution systems.
Accurate short-term solar irradiance forecasting is necessary for the effective grid integration of photovoltaic systems. This study proposes a novel two-stage framework, VMD-TFT, which integrates the Variational Mode Decomposition (VMD) with the Temporal Fusion Transformer (TFT) for multi-horizon forecasting of direct normal irradiance (DNI). In the first stage, the input signal is decomposed into intrinsic mode functions (IMFs) using VMD, enabling the separation of multi-scale temporal patterns. In the second stage, each IMF is modeled using a dedicated TFT sub-network, and the final prediction is obtained through reconstruction of all components. The proposed approach is evaluated on two benchmark datasets (Folsom, California, and NREL SRRL, Colorado) across multiple forecasting horizons (1h, 3h, and 6h). Experimental results demonstrate that VMD-TFT consistently outperforms six baseline models, including statistical, machine learning, and deep learning approaches. For the 1-hour horizon, the model achieves an R² of 0.991 and an RMSE of 12.5 W/m², showing significant improvement over the vanilla TFT baseline. The performance gain becomes more pronounced at longer horizons, with up to 20.3% RMSE reduction at 6 hours. In addition to point forecasting, the framework provides probabilistic predictions via quantile regression and demonstrates robustness through cross-site validation. The results indicate that integrating signal decomposition with attention-based deep learning improves both forecasting accuracy and model stability for complex, non-stationary solar irradiance signals. On the Folsom benchmark, the proposed VMD-TFT yields R2 values of 0.991, 0.972, and 0.948, and RMSE values of 12.5, 28.3, and 42.1 W/m², respectively, across the investigated horizons. The framework achieves a 29.8% reduction in 1-hour RMSE over the vanilla TFT baseline, demonstrating statistical significance under the Diebold-Mariano test (p < 0.001).
Gestational diabetes mellitus affects approximately 14% of pregnancies worldwide. Especially in Thailand, the prevalence reaches 29.2%, among the highest in Southeast Asia, because access to personalized nutritional counseling remains limited, particularly in Thai-language healthcare settings. This study developed and validated a Thai-language multi-agent conversational platform comprising a Passive Agent for trimester-aware dietary and health information retrieval and an Active Agent for proactive dietary monitoring and glucose management through multi-turn interactions. The system is aimed to support patient education and self-management. We validated our system through task-specific assessments. The Passive Agent was assessed by ten healthcare professionals across nutrition, gestational diabetes mellitus, and antenatal care domains using accuracy, source attribution, Global Quality Score, trimester-specific appropriateness, response repeatability, and inter-rater reliability. The Active Agent was validated on 125 nutritionist-curated food items for carbohydrate estimation accuracy and consistency against nutritionist-established ground truth. Passive Agent achieved the highest accuracy in gestational diabetes mellitus (88–96%), followed by nutrition (88%) and general antenatal care (68–72%), with question-level accuracy of 84%. Global Quality Score and trimester-specific appropriateness were highest in nutrition, and response repeatability exceeded the threshold (mean cosine similarity 0.855). Active Agent achieved 92.8% carbohydrate estimation accuracy within 0.5 exchanges, a mean absolute error of 0.107 ± 0.234 exchanges, and a coefficient of variation below 10% for 98.4% of items, with no significant difference from nutritionist ground truth. These findings demonstrate the feasibility and expert-rated quality of a RAG-based agentic conversational prototype, supporting future testing with real patients in full RCT experiments before consideration as a national maternal healthcare tool.
This paper presents a novel geometric configuration allowing the monolithic and compact manufacturing of a very wide-angle coverage antenna for radar applications. Geodesic lenses are able to provide a large scanning range while being efficient owing to their metallic low-loss topology. Two superimposed lenses are combined with a rotation to reach an upgraded field-of-view of 220° in Ku-band. The number of beams and the rotation angle are carefully chosen to minimize the scan loss and get non-overlapping patterns. The full-metal device is printed thanks to additive manufacturing for maximal performance and compactness through the assembly-free structure. The fabrication complexity is increased in comparison with the literature due to the rotation between the lenses. Hexagonal waveguide feeds are introduced with a symmetrical excitation strategy, alongside a lens profile modification, to control overhang angles and enable the optimal realization of such a concept. A very good match between the simulations and the measurements is attained thanks to the electromagnetic and fabrication co-design. The prototype measured reflection coefficients are below -10 dB between 12.7 and 22.8 GHz. The measured coverage is 224° at 13.5 GHz with a gain of 14.4 dBi and a scan loss of 3.1 dB.
