Sensors doi: 10.3390/s26217037

Authors: Katarzyna Bobkowska Agata Kaczyńska-Miotk Adam Kosiński

The right ventricular outflow tract (RVOT) is an anatomically complex and spatially variable cardiac region that is difficult to characterise using isolated manual measurements. Although three-dimensional reconstruction techniques enable detailed visualisation of RVOT morphology, reproducible workflows for quantitative, level-based analysis of RVOT geometry remain limited. The aim of this methodological study was to develop and evaluate a semi-automated workflow for quantitative analysis of RVOT geometry based on 3D point-cloud models. This study represents a secondary methodological analysis of an existing anatomical dataset, introducing a new semi-automated workflow for systematic cross-section generation and quantitative descriptor extraction. Anatomical RVOT models were prepared from post-mortem material using silicone casting and digital photogrammetry. After point-cloud preparation, primary cross-sections were used to estimate a global directional reference. A mathematically defined inferior boundary plane and a predefined superior anatomical limit were then used to generate secondary cross-sections along the standardised RVOT segment. For each retained section, positional, ellipse-based, distance-based and area-based descriptors were extracted. The workflow was applied to 60 RVOT point-cloud models. After quality control, 661 secondary cross-sections were retained for analysis, with 5–16 sections per model and a mean of 11.02 ± 2.47 sections per model. Methodological validation included synthetic testing, comparison with manual measurements, and inter-observer assessment. Agreement with manual dimensions was high (ICC = 0.905–0.937), while inter-observer landmark variability was low (RMSE = 0.41 mm). An additional exploratory inter-observer assessment of three RVOT models yielded mean model-specific RMSE values of 12.97 mm2 for cross-sectional area and 0.54 mm for both ellipse axes. The proposed workflow provides a predefined geometric framework for quantitative description of RVOT geometry and should be regarded as a methodological proof of concept rather than a clinically validated measurement system.

Sensors doi: 10.3390/s26217036

Authors: Shuyi Gao Ren Jin Defu Lin Weiqiang Xiao Jinmeng Chen

Overtaking planning and decision-making are central to performance in autonomous racing. Limited maneuvering space on narrow closed circuits makes high-speed overtaking difficult and increases the risk of failed overtaking attempts and collisions. To address this problem, we formulate overtaking as robust tactical feasible-set selection and develop a Stackelberg-supervised policy that translates response-set reasoning into persistent tactical corridors for constrained planning and closed-loop control. First, a robust Stackelberg response-set teacher provides action-prior and action-conditioned robust-value supervision to a Truncated Quantile Critics (TQC) policy that takes sensor-derived opponent-state estimates as input and learns passing-side intention, corridor parameters, and speed demand. Second, a hysteretic finite-state machine (FSM) preserves the selected passing side, while a deterministic corridor carver combines the learned corridor parameters with the evolving track and interaction geometry to construct a time-varying Frenet corridor. Finally, a corridor-constrained sequential second-order cone programming (SOCP) planner and a speed profiler generate a continuous path–speed reference, which is tracked using reference-conditioned contouring-error nonlinear model predictive control (RC-NMPC) to produce closed-loop vehicle commands. Hardware-in-the-loop experiments were conducted using an Emirates Autonomous Vehicle (EAV24) formula race-car simulator with Gaussian noise and packet drops applied to localization and perception outputs. Across 1000 laps, the method achieved a lead-acquisition rate of 94.7% and a mean time to lead acquisition of 4.82 s. Compared with three baselines using the same execution stack, the lead-acquisition rate was 31.2–66.2 percentage points higher and the mean time to lead acquisition was 3.01–6.41 s shorter; no collisions or track-boundary violations were observed.

Sensors doi: 10.3390/s26217035

Authors: Ya-Zhou Zhou Le-Tian Fang Zhen-Nan Zhou Nan Zhou Yun Wang Fei Shan

Resistive switching devices are attractive candidates for nonvolatile memory and sensory information storage in intelligent sensor nodes, but their use near sensors remains limited by switching variability and resistance state instability. In this work, a two-dimensional MoS2 interlayer is introduced into an Ag/HfO2/TiN memristor stack to form a MoS2/HfO2 heterostructure device, and its electrical characteristics are compared with an Ag/HfO2/TiN control device fabricated under the same oxide deposition and annealing conditions. The heterostructure device exhibits well-defined and reproducible bipolar resistive switching with an on/off ratio of up to approximately 103, together with markedly improved cycle-to-cycle overlap of the current-voltage characteristics, whereas the switching window of the control device narrows and eventually closes after repeated sweeps. X-ray photoelectron spectroscopy of the O 1s core level shows that the oxygen-vacancy-related component decreases from 18.31% to 5.12% of the fitted spectral envelope after introduction of the MoS2 interlayer, while the lattice-oxygen component remains essentially unchanged. This selective reduction is attributed to interfacial passivation of electrically active vacancies rather than to bulk oxygen extraction, and it cannot be explained by signal attenuation of the buried oxide under the MoS2 overlayer. Retention measurements show that the heterostructure device maintains stable high- and low-resistance states over the measured 103 s window, while the high-resistance state of the control device decays by orders of magnitude within the same period. The switching polarity, the correlation between vacancy chemistry and switching uniformity, and the stability of the ruptured-filament state are consistent with an interface-guided valence-change mechanism, although a minor contribution of Ag-ion migration cannot be excluded. These results indicate that MoS2/HfO2 heterojunction engineering is an effective route toward mitigating variability and retention degradation in oxide-based memristors for intelligent sensing hardware. The measured resistance window, repeatable switching, and state retention are relevant to local storage of sensor features, calibration coefficients, and adaptive thresholds in robotic perception. The present work establishes the memory component of such sensor nodes and does not claim direct transduction or a complete robotic sensing system.

Sensors doi: 10.3390/s26217034

Authors: Fernando Palacios Ana Redondo Rodriguez Hélder Rodríguez Salvador Domínguez-Durante Daniel Sanz-Merodio Miguel López

Detecting people from unmanned aerial vehicles (UAVs) in coastal environments is challenging because targets are often small, distant, and observed against complex backgrounds. This case study presents an application-driven system-level comparison of frame- and event-based sensing vision systems. We developed a purpose-built multimodal UAV dataset of beach users, bathers, and surfers using frame and event-based cameras. Frame images and events were evaluated using feasible standard and converted spiking YOLOv2-derived configurations on a Jetson Orin Nano and a BrainChip AKD1000 neuromorphic processor. The experimental setup serves as a challenging stress test to expose limitations and relative behavior across configurations. The evaluation covers end-to-end processing performance and energy and sensor power. Frame-based YOLOv2 achieved the highest detection performance, with an F1-score of 0.215 and an mAP50 of 0.212. The best spiking configuration, sYOLOv2 with event input, achieved an F1-score of 0.172 and an mAP50 of 0.114. Frame-based YOLOv2 on the Jetson GPU also achieved the lowest end-to-end processing energy cost (2886.6 mJ per sample) and highest throughput (2.011 samples/s). At the component level, the event camera reduced operating power by 82.1% relative to the frame camera, while the AKD1000 consumed approximately 1.3 W during backend execution. However, these component-level savings did not yield an end-to-end energy advantage because of host-processing overhead and lower throughput. The results highlight the need to assess sensing and complete-pipeline costs when evaluating the efficiency of neuromorphic embedded vision systems.

Sensors doi: 10.3390/s26217032

Authors: Nadeem Fareed Izaya Numata Carlos Alberto Silva Susan J. Prichard

Ground-point classification (GPC) is essential for light detection and ranging (LiDAR)-derived terrain and vegetation products, yet most of the well-established GPC algorithms often require parameter calibration across varying terrain, vegetation, point densities, and sensor modalities—airborne laser scanning (ALS), unoccupied aerial vehicle laser scanning (ULS), mobile laser scanning (MLS), and terrestrial laser scanning (TLS). This study introduces FAST-GC, a fully adaptive, self-tuning, parameter-free GPC algorithm designed for ALS, ULS, MLS, and TLS LiDAR sensors. FAST-GC was evaluated against cloth simulation filtering (CSF), progressive TIN densification (PTD), and progressive morphological filtering (PMF) across benchmark sites (n = 6), with additional terrain-complexity, slope, and sampling-density assessments. FAST-GC maintained high ground completeness with limited commission errors, particularly under complex topography. At slopes of 70–80°, the FAST-GC-derived digital terrain model (DTM) mean absolute error (MAE) was 1.00 m, compared with 3.12–12.75 m for the reference counterparts—CSF, PMF, and PTD. Across 20 ALS sites, reducing median point density from 9.84 to 3.84 points/m2 maintained F1 ≈ 0.97, intersection over union (IoU) ≈ 0.95, and Matthews correlation coefficient (MCC) ≈ 0.85, while DTM MAE increased slightly from 0.07 to 0.11 m. FAST-GC-derived DTM (0.5 m) from concurrent ALS, ULS, and TLS showed strong cross-platform agreement (R2 and concordance correlation coefficient (CCC) ≥ 0.99; RMSE = 0.10–0.16 m). Processing 382 ALS tiles (~23 km2) further demonstrated scalable execution and automated terrain-product generation. These findings establish FAST-GC as a reproducible, sensor-aware framework that preserves terrain morphology across heterogeneous LiDAR datasets without scene-specific parameter calibration.

Sensors doi: 10.3390/s26217033

Authors: Fengxiang Wang Genghua Huang Xiangfeng Liu Rong Shu

Discrepancies between geometric priors and LiDAR observations can limit pose tracking of non-cooperative targets. This paper proposes online reference template adaptation that incorporates observations selected through registration-quality, incident-angle, and temporal-support criteria into a CAD-derived template. Two numerical simulations with known pose ground truth and two ground-based experiments evaluate the method under target rotation and close-range approach. In the approaching simulation, online adaptation reduces translation and rotation RMS errors from 1.44 to 0.52 cm and from 1.10∘ to 0.40∘, respectively. In the rotating simulation, the corresponding errors decrease from 4.61 to 4.37 cm and from 3.97∘ to 3.32∘. In the ground-based rotational experiment, online adaptation achieves a mean registration RMSE of 3.29cm, approximately 31% lower than the fixed CAD baseline and close to the offline reconstructed baseline (3.05cm), while reducing threshold exceedances from 32 to four isolated frames. In the close-range experiment, roll estimates remain comparatively stable as the visible region shrinks, whereas the fixed CAD baseline exhibits more than 10∘ of roll drift. These results support improved pose accuracy in the tested simulations and greater tracking stability in the ground-based experiments, without dedicated prior multi-view reconstruction.

Sensors doi: 10.3390/s26217031

Authors: Changjun Li Jinjun Cheng Haizhen Zhu Bincheng Wen Ziyi Yu

Modernized equipment features a multi-component structure, which is characterized by complex degradation behaviors, noisy monitoring measurements, and considerable uncertainty in life prediction, making accurate remaining useful life (RUL) forecasting a challenging task. To address this issue, this paper takes an aero-engine as a typical research subject and proposes a multi-component systems RUL prediction method that integrates physical priors and a variational spatiotemporal graph convolutional network. Firstly, graph-structured data are constructed according to the physical topological relationships among multiple engine components, and an initial adjacency matrix is generated to endow the model with prior physical knowledge of sensor dependencies. Secondly, a spatiotemporal feature extraction module is established, which combines graph convolution, dual attention mechanisms, and temporal convolution to adaptively select informative sensor channels and critical spatiotemporal features. Finally, a variational information bottleneck mechanism is introduced. By means of mask-based noise reduction and dynamically learned probabilistic graph structures, information relevant to life-related characteristics is screened, and the processed features are fed into a multi-layer perceptron to implement RUL regression prediction. Validations on the public C-MAPSS aero-engine dataset demonstrate that the proposed method adequately combines the merits of physical priors and data-driven paradigms. It effectively mitigates noise interference and reduces prediction uncertainty, and can provide technical support for health management and life prediction of multi-component aero-engine systems.

Sensors doi: 10.3390/s26217030

Authors: Marcel Reith-Braun Felix Kronenwett Preetham Suresh Georg Maier Harald Kruggel-Emden Thomas Längle Jürgen Beyerer Florian Pfaff Uwe D. Hanebeck

Pneumatic ejection has dominated sensor-based sorting for decades, yet diffuse air jets cause bycatch at high mass throughput, and compressed-air generation accounts for a large share of operating cost. We propose (i) a grid-based sorting architecture replacing the conventional single nozzle bar with a two-dimensional grid, and (ii) impact-based mechanical actuators within the grid. Together, these eliminate compressed air entirely while increasing spatio-temporal resolution by leveraging redundant particle–actuator assignments. To coordinate online ejection—deciding which actuators to trigger and when—we formulate sorting as a finite-horizon soc problem and introduce mtcr. mtcr integrates olf soc, mixed discrete–continuous-time modeling, situation-aware action-space reduction, and a greedy rollout heuristic, supported by analytically tractable evaluation of expected sorting costs. In high-mass-flow simulations with equal accept/reject ratios, mtcr maintains both tnr and tpr above 90 , demonstrating effective actuator coordination and bycatch reduction. Laboratory experiments on a 56-solenoid prototype show that the grid significantly exceeds a state-of-the-art pneumatic chute sorter in rejection performance at matched bycatch level, with actuation energy being at least 79 times lower. Grid-based mechanical sorting thus offers a promising route toward precise, energy-efficient separation for cost-sensitive industries such as recycling.

Sensors doi: 10.3390/s26217029

Authors: Juliette Ngan Yuanwei Liu Ali Asgar Saleem Bhagat

Wearable gait assessment using foot-mounted inertial measurement units (IMUs) can extend mobility evaluation beyond the laboratory, but performance is limited by integration drift, magnetic disturbances, and gait event detection sensitivity during outdoor overground walking. This study presents an insole-embedded nine-axis IMU pipeline combining orientation estimation using the Versatile Quaternion Filter with zero-velocity update-aided error-state Kalman filtering and Rauch–Tung–Striebel smoothing. Twenty-one healthy adults performed slow, normal, and fast 100 m straight path walks and normal-speed 437.7 m full-lap walks on an outdoor track. Seventy-five retained trials were compared with the Movella MVN Awinda reference using task-specific absolute-agreement intraclass correlation coefficients (ICCs) and mixed-effects Bland–Altman analysis accounting for repeated observations within participants. Agreement was excellent for stride length and stride time (ICCs: 0.928–0.961 and 0.989–0.997, respectively), good to excellent for gait velocity (0.894–0.957), moderate to excellent for stance time (0.679–0.942), and moderate to good for swing time (0.544–0.790). Mean biases were +0.036 m/s for gait velocity and +1.54 cm for stride length; temporal biases ranged from −0.005 to +0.005 s. Walking distance mean absolute percentage errors ranged from 0.55% to 1.20%. These findings support the feasibility of insole-based spatiotemporal gait assessment during predefined outdoor walking in healthy adults. Further validation against independent reference systems and in clinical populations and more complex environments is needed.

Sensors doi: 10.3390/s26217028

Authors: Huanjun Chen Yan Wang Jingran Sheng Xinfu Pang Zedong Zheng

Reliable short-term wind power forecasts are difficult to obtain because power records are strongly non-stationary and hybrid networks are highly sensitive to parameter choices. This study proposes a Pearson–HOA–VMD–CNN–iTransformer–LSTM framework to address these challenges. Pearson coefficients are used to screen the candidate meteorological inputs. The hiking optimization algorithm (HOA) subsequently searches the mode number and penalty parameter of variational mode decomposition (VMD), together with the principal hyperparameters of the forecasting network. The optimized decomposition reduces sequence complexity, while a convolutional neural network (CNN) extracts local fluctuation patterns. iTransformer captures global variable dependencies using variate-wise self-attention, and long short-term memory (LSTM) models learn long-term temporal information. The framework was evaluated using four datasets containing 2976, 2880, 2976, and 2976 records, respectively, drawn from 2020 operational data from two Chinese wind farms with installed capacities of 99 and 200 MW. The sampling interval was 15 min, and the forecast horizons were 15, 30, and 45 min. Tests on four datasets show consistent gains over CNN–iTransformer–LSTM. Averaged across the datasets at the 15-min forecast horizon, RMSE, MAE, and MAPE decrease by 67.48%, 71.17%, and 49.66%, respectively, and mean R2 rises from 0.9672 to 0.9974.

Sensors doi: 10.3390/s26217027

Authors: Yuanbo Ran Haijiang Wang Taofeng Gu Fang Liu Bo Zhang

As a prevalent meteorological hazard across China, severe thunderstorm winds originate directly from storm cells. With the dense networking of Doppler weather radars in recent years, networked radar observations have become an essential technical support for the operational monitoring and early warning of severe thunderstorm winds. Traditional storm cell identification, tracking and nowcasting methods are mainly tailored to single-radar datasets. To overcome the limitations of traditional methods, this study proposes a novel method for storm cell identification, tracking, and nowcasting based on networked radar data. The method comprises three steps: identifying 3D storm cells via digital image processing and assessment their intensity using fuzzy logic; constructing a minimum cost function for spatiotemporal matching and continuous tracking; and finally, adopting the unscented Kalman filter and the adaptive damped Holt method to forecast the future trajectory and intensity, respectively. Compared with the traditional operational nowcasting methods adopted by meteorological services in China, namely storm cell identification and tracking (SCIT) and thunderstorm identification, tracking, analysis and nowcasting (TITAN), the proposed method exhibits growing advantages in storm location prediction within the 60-min forecast lead time, with controllable decay of probability of detection (POD) and critical success index (CSI) and low false alarm rate (FAR). The developed intensity forecast product enriches the storm forecasting framework, with its mean absolute error (MAE) kept within a controllable range to support operational short-range nowcasting.

Sensors doi: 10.3390/s26217026

Authors: Pierre Boulanger

MedROAD V3 is a real-time FHIR R4 streaming platform that turns clinical sensor streams into calibrated deterioration risk, evaluated retrospectively on a cardiac ICU cohort. FHIR event sourcing, Kafka and Faust assemble an 86-element feature vector every five minutes. An XGBoost, LSTM and Transformer ensemble, combined by logistic meta-learning with SHAP attribution, scores each window; a write-back client returns the alert to the chart with a runtime-filtered language-model narrative. On 5049 held-out MIMIC-IV patients it reaches AUROC 0.816 at an Expected Calibration Error of 0.001, with a median inference path of 3.1 ms, so latency is governed by the EHR rather than by computation. The FHIR Observation resource is device-agnostic, so the pipeline ingests wearable streams unmodified. Three limits qualify these results. It is retrospective, and the remote-monitoring figures are a feature-restriction analysis on ICU data, not a wearable validation. Twenty-nine columns recording only how a patient was documented and managed, none physiological, reach AUROC 0.767 on the same split, so much of the discrimination reflects care process rather than physiology, although removing them costs only 0.007. Because OpenEMR 7.0.2 reference server grants no clinical FHIR write scope, the loop closes over an alternative path using its REST API, writing the alert as a Subjective-Objective-Assessment-Plan (SOAP) note on the encounter. The substitution works but is vendor-specific rather than portable, so the closed loop shown here is conditional on the target server.

Sensors doi: 10.3390/s26217024

Authors: Kang In Choi Eun Hak Lee Seo Jeong Kim Chang Ho Yu Youngsun Hong Sunghwan Jeong

This study develops a deep multimodal learning framework for lane-level traffic-speed estimation using five sensing sources: vehicle detection systems (VDS), closed-circuit television (CCTV), light detection and ranging (LiDAR), remote traffic microwave sensors (RTMS), and automatic vehicle classification (AVC). The framework integrates sensor-specific features, road attributes, modality availability, and spatial relationships. Using traffic data from Siheung, Republic of Korea, the proposed model was evaluated against five single-modal baselines. The multimodal model achieved a weighted mean absolute percentage error (WMAPE) below 10% across evaluated time steps, outperforming single-modal models in stability and cross-modal transferability. Improvements were more pronounced in selected multi-lane environments and areas with limited sensing coverage. In route guidance, lane-level multimodal information reduced mean minimum travel time by 7.31% compared with link-level information. These findings demonstrate the potential of multimodal integration for accurate lane-level estimation, recovery of missing sensor observations, and improved route guidance.

Sensors doi: 10.3390/s26217025

Authors: Pal Varga

The Eclipse Arrowhead Framework provides a service-oriented architecture for distributed industrial automation through service discovery, authorization, orchestration, and quality-of-service-aware service selection. While the framework formally defines the execution of individual service interactions, it does not provide a mathematical semantics for coordinating multiple dependent service invocations into executable workflows. This paper presents a compact mathematical execution model for Arrowhead workflow choreography. Rather than introducing a new orchestration mechanism, the proposed model extends the Arrowhead Framework conservatively by treating workflow execution as the repeated compilation of workflow steps into ordinary Arrowhead service requests. The model introduces formal representations of workflow recipes and workflow sessions, together with a compact execution model that maps executable workflow steps to ordinary Arrowhead orchestration requests and service invocations. A worked example demonstrates how the notation can be instantiated for a generic setting, whereas another example shows a concrete temperature-control workflow, and a third example a shared-tool delivery workflow in a forklift assembly plant. The resulting formalization preserves the service-oriented semantics of the Arrowhead Framework while providing a mathematically precise model for workflow execution, including measurement workflows that coordinate sensing, validation and actuation services, in dynamic industrial environments.

Sensors doi: 10.3390/s26217023

Authors: Saman Iftikhar Ammar Rafiq Syed Rizwan Hassan

Internet of Things (IoT) and Internet of Medical Things (IoMT) sensor deployments are pushing security decision-making from centralised servers towards resource-constrained edge gateways, where heavyweight deep learning detectors exceed the available compute, memory, and latency budgets, and where opaque alarms erode operator trust. This paper proposes ExLightNet, a lightweight one-dimensional convolutional neural network that couples depthwise-separable convolutions and squeeze-and-excitation channel attention with a tempered class-balanced training recipe (focal loss plus f−1 sampling) and per-channel INT8 weight quantisation, and attaches SHAP attributions directly to the deployed model. We evaluate the framework on the public RT-IoT2022 benchmark of 123,117 real bidirectional flows from smart-home and sensor IoT devices spanning nine attack behaviours and three normal device classes, an environment representative of the auxiliary sensing layer of home-based care settings. ExLightNet attains an attack-class recall (sensitivity) of 0.9976, missing only 40 of 16,592 attack flows, with macro-F1 0.9860 and AUC 0.9985 for binary attack detection (the 99.49% accuracy is retained for comparability with prior work, but accuracy is a misleading headline metric under the severe class imbalance of network traffic), and a macro-recall of 0.863 (97.70% accuracy, macro-F1 0.8621) across twelve fine-grained classes, surpassing a 4.5× larger standard CNN on the binary task while remaining within 0.0021 macro-F1 of Random Forest. Post-quantisation, the model shrinks from 162 KB to 37 KB (4.4×) with a macro-F1 drop of at most 0.007, sustaining 0.50 ms single-flow CPU latency, 66× faster than an accuracy-equivalent Random Forest. On a Raspberry Pi 4 gateway, the INT8 model sustains 2197 flow decisions per second across four Cortex-A72 cores, 21.4× faster than an accuracy-equivalent Random Forest, confirming the deployment profile on measured gateway-class hardware. SHAP analysis shows that forward payload statistics, backward header-size extremes, and packet-rate features dominate decisions, yielding operator-readable rationales.

Sensors doi: 10.3390/s26217022

Authors: Xiaohan Wang Chunlei Song Yuning Jiao Guoxiang Rao Jianhua Xu

Hybrid inertial navigation systems (HINSs) integrate physical, stabilized platforms, strapdown attitude algorithms, and rotation modulation technology, offering unique performance advantages. However, the non-ideal orthogonality of the gimbal frames in the rotation mechanism introduces frame non-orthogonality errors, which limit the accuracy of high-precision attitude extraction and transfer alignment. To address the limitations of existing calibration methods—such as incomplete modeling, poor noise immunity, or dependence on the initial mounting attitude—this paper establishes a complete frame non-orthogonality error model and proposes two improved calibration methods: a three-position self-calibration method and a recursive least squares (RLSs) calibration method. Monte Carlo simulations show that the three-position method is simple in principle but sensitive to sensor measurement noise (accelerometer noise and gyroscope angular random walk), whereas the RLS method recursively fuses observation information and reduces the average standard deviation of the estimated non-orthogonality angles to 0.044″, with superior stability and repeatability. On this basis, the effect of non-orthogonality angle compensation on transfer alignment accuracy is verified, showing that the improvement of horizontal misalignment angles exceeds 96%. Finally, the proposed methods are validated through physical experiments on a HINS. After compensation with the RLS method, the peak residual roll attitude error decreases from 18.53″ to 10.6″, effectively improving the overall navigation performance of the system.

Sensors doi: 10.3390/s26217021

Authors: Yuhang Li Chong Zhang Wei Huang Qingtian Lü

Airborne electromagnetic (AEM) measurements are vulnerable to mixed environmental noise, and rapid signal decay hampers weak transient recovery. We propose WL-CNN-BiLSTM, integrating dual-stage wavelet filtering, parallel multi-scale convolutions, residual skip connections, bidirectional recurrent units, and a hybrid reconstruction objective for complex TEM waveforms. On the archived 20,000-trace benchmark’s 3000-trace test subset, it achieved an aggregate physical-domain SNR 17.53 dB (RMSE 1.48 × 10−10 T; PCC 0.989), versus 11.49 dB for PC-BiLSTMNet, 11.76 dB for validation-selected Median filtering, and 13.04 dB for VMD; the per-trace SNR P10 was −3.41 dB (minimum −25.45 dB). These archived results characterize stored numerical targets, not independently verified AEM responses. On an independent 1000-trace blind test, a separately trained layered-earth configuration using training-only scalers achieved a median SNR of 5.91 dB versus 0.39 dB (Median) and −0.05 dB (VMD), but poorer lower-tail SNR (P10 −43.93 dB) and a median of NRMSE 7.402 over 101–1001 μs. Proprietary iFTEM-II windows showed waveform changes versus the engineering curve, but unavailable proxy provenance/calibration made this qualitative evidence, not field-accuracy validation. A matched p = 0 comparison, three-seed ablations, complete training-only-scaler retraining, and out-of-distribution diagnostics delineate distributional dependencies; temporal feature weighting is an implemented design component, not a demonstrated late-time enhancer. It shows aggregate reconstruction gains within the benchmark and characterizes dataset- and protocol-dependent limitations.

Sensors doi: 10.3390/s26217020

Authors: Jiaoyang Li Zhao Liu Xingliang Zhao Guanyu Piao

Moisture-induced deformation, cracking, and biological deterioration threaten wooden cultural heritage, making reliable, nondestructive moisture assessment essential for preventive conservation. The microwave method has been recognized as an emerging method with advantages of noncontact, fast and high accuracy. However, conventional microwave transmission-based moisture detection methods require access to both sides of a specimen, hindering the inspection of fixed components with inaccessible rear surfaces. This study investigates single-sided, non-contact wood moisture assessment using the principle of free-space microwave reflection. A three-dimensional full-wave antenna–wood model identified significant moisture-dependent variations in a valley near 1.1 GHz and its adjacent peak. Their magnitude difference was defined as a valley-depth feature, with an optimized 100 mm standoff selected to balance linearity and sensitivity. The sensing system was developed based on the optimized experimental parameters from the simulation. And the calibration test showed a linear relationship between valley depth and relative permittivity (R2 = 0.98). Pine and paulownia specimens below 30% moisture content were measured in two polarization orientations, yielding short-term repeatability standard deviations less than 0.013 dB. Independent validation achieved the lowest errors for paulownia with the electric field perpendicular to the grain, with RMSE = 1.27% and MAE = 0.99%. An in-situ trial on a historic wooden door demonstrated field deployment feasibility and measurable spatial and temporal variations in the reflection response. These findings support the potential of the method for single-sided, non-contact wood moisture assessment.

Sensors doi: 10.3390/s26217019

Authors: Zanhao Liang Ruixiao Huang Xuhui Ma Binbing Liu

Low-light image enhancement (LLIE) is fundamentally a restoration problem rather than a brightness adjustment task, as it aims to recover the underlying scene information from images captured under low signal-to-noise ratio conditions. However, most existing methods adopt globally uniform enhancement schemes and overlook the degradation discrepancies across semantic regions, limiting overall performance. Recent semantic-guided LLIE methods often neglect how low-light degradation impairs the fidelity of semantic priors, preventing them from exploiting semantic information. To address these issues, we propose an implicit semantic guidance branch that employs a Light-up Module and a Feature Alignment Module to obtain accurate semantic features and adaptively integrate them into the feature representation space. Furthermore, we introduce an explicit semantic guidance branch that constructs a Semantic-Guided Pixel Consistency Loss using accurate semantic maps obtained by SAM 3, mitigating segmentation errors caused by cross-dataset label space discrepancies. Additionally, we develop a carefully designed low-light degradation model to synthesize training data and adopt a pre-training strategy, enabling the network to learn fine-grained semantic-aware brightness mappings. Based on these components, we develop a hybrid explicit–implicit semantic guidance framework. Extensive experiments demonstrate that our method outperforms state-of-the-art approaches in image quality, noise suppression, and structure preservation.

Sensors doi: 10.3390/s26217018

Authors: Mohammed Almukaynizi Abdulaziz Alajaji

Public Snap Map activity, a short-lived stream of stories pinned to places, can serve as an opportunistic urban sensor, and this study measures what it reports and what it requires to report reliably. From more than 1.18 million public snaps collected on a 1 km grid across three contrasting Saudi cities in 2024, activity is highly concentrated, to degrees that differ by city. Over the full windows, Makkah leads on every measure (Gini 0.71 [0.68, 0.73]), with one Grand Mosque tile holding 2.7% of its snaps, and its lead survives a spatial-block bootstrap. Riyadh (0.59 [0.57, 0.60]) and Dammam (0.56 [0.54, 0.59]) are close in degree and differ in shape: Dammam concentrates in a strong head of busy tiles, Riyadh in a long active tail. Over the week that all three windows share, Makkah does not lead, because its estimate rises with accumulated observation. Two operating requirements follow. Concentration estimates stabilized after 23 to 27 days in the two campaigns observed long enough to confirm it and after 12 to 14 days once collection start-up had passed. Under binomial thinning with item-independent visibility, the Gini is biased downward, Makkah’s lead survives at 5% retention, and the Riyadh–Dammam ordering inverts at 10%.

Sensors doi: 10.3390/s26217017

Authors: Chaimae Saadouni Saad El Jaouhari Nouredine Tamani Lina Mroueh Soumia Ziti

Bluetooth Low Energy (BLE) devices are ubiquitous across consumer, industrial, and healthcare IoT ecosystems, yet their resource-constrained nature typically precludes conventional cryptographic authentication, leaving device identification as a security primitive for device management, access control, and threat detection in IoT environments. Current BLE fingerprinting techniques are restricted by rigid cryptographic hashes, machine learning models that require periodic retraining to cover new devices, or single-layer feature sets vulnerable to missing data. In response to these limitations, we introduce a Nilsimsa-based identification approach that performs BLE device type/manufacturer identification through merging link-layer features and advertised GATT service information before network access is granted to cope with missing or inaccessible data. We analyzed various combinations of features from different layers to assess how merging them impacts identification accuracy. The exhaustive analysis uses 10 test objects and 10 reference hashes. The large-scale evaluation uses 5000 test records and 10,000 reference hashes covering 104 device type–manufacturer pairs. Key findings highlight that cross-layer features are important to cope with missing features. The minimum subset size was found at k = 10 with a subset-coverage criterion: at least 95% of all subsets of 10 features reach an accuracy of 90% or more. Our proposal not only facilitates the seamless enrollment of new devices through reference-hash registration without retraining but also enables confidence-based anomaly detection using its similarity score, providing an incrementally extensible approach for BLE security monitoring. Overall, our approach advances BLE fingerprinting by introducing a schema-agnostic, retraining-free, cross-layer identification model tailored for the dynamic environment of IoT deployments.

Sensors doi: 10.3390/s26217016

Authors: Kishor Arvind Mohandas Sivakumar Laurent Segers Bruno da Silva Argha Sen Sujay Narayana R. R. Venkatesha Prasad

Low-cost multi-modal sensors make high-resolution exposomic and urban-health monitoring affordable, but their data are only as useful as they are trustworthy. Measurement bias, cross-sensitivity, temporal drift and domain shift between devices and sites erode reliability, and reference instruments are rarely available for every sensor. We present a two-part reliability workflow for this realistic setting, in which reference data exist for only a subset of sensors, and evaluate it on five custom multi-modal devices deployed at two Brussels sites (9 to 68 days of data per device). For calibration, we benchmark Multiple Linear Regression, Elastic Net, Random Forest, Gradient Boosting and a Domain Adversarial Neural Network (DANN) for O3, NO2, PM2.5 and PM10 on the source device and on other devices at the same site and at the other site using MAE, RMSE, R2, bias, device-balanced aggregation and block-bootstrap intervals. Tree ensembles achieve the lowest source-device errors for most pollutants, yet for PM10, they lose their advantage over the uncalibrated sensor when transferred, a caution for any deployment that calibrates on one unit. DANN attains the highest mean Calibration Transfer Ratio in the transfer settings, although with the available device pairs, the differences are not statistically significant. For modalities without a reference, a cross-sensor residual monitor combined with statistical detectors reliably flags injected gradual drift in sensors that can be reconstructed from the others (UV bands, PM1, temperature, humidity and CO), while Lux, pressure and VOC remain harder. The monitor detects cross-sensor inconsistency rather than isolating the faulty sensor, and the results reflect an injected fault model, not real sensor ageing. Together, these results give a transparent, reproducible baseline and clear practical guidance for building trustworthy low-cost sensor networks.

Sensors doi: 10.3390/s26217015

Authors: Ibrahim Mutambik

Smart-city wireless sensor and Internet of Things (IoT) networks are progressively transitioning toward sustainable energy sources. During this transition, rechargeable energy-harvesting nodes must coexist with conventional devices powered by non-rechargeable batteries, creating partially sustainable IoT networks with conflicting energy-management objectives. Energy-harvesting nodes seek energy-neutral operation, whereas conventional nodes must minimize consumption to extend their operational lifetime. Coordinating these devices requires the exchange of energy information (EI), particularly battery state of charge. However, disclosing such information can enable adversaries to identify energy-constrained nodes and target them through energy-depletion attacks. This study examines the existing Dual Energy PROfile for interNet-of-thiNgs with Enhanced Security (DEPRONN-ES) framework for heterogeneous smart-city IoT networks. The framework integrates dual-profile energy management, energy-aware routing, adaptive data aggregation, and differential privacy. The focus is their coordinated application to heterogeneous energy resources and privacy-preserving EI exchange, rather than the introduction of new individual techniques. The framework was implemented using Contiki-NG and evaluated through Renode emulation with real indoor light-intensity traces from Columbia University to model energy harvesting by rechargeable nodes. The results show that, with a privacy budget of 0.1, DEPRONN-ES reduces the attacker’s target-identification success rate from 100% to 13.33%, representing an 86.67-percentage-point reduction. This reduction concerns targeting accuracy and does not imply prevention of malicious traffic. It also enables sustainable and conventional nodes to operate according to their distinct energy objectives. The reported findings suggest potential benefits from coordinating privacy protection and energy management under the evaluated conditions.

Sensors doi: 10.3390/s26217014

Authors: Dong Zhou Zhe Huang Haiyang Li Mengyun Cao Chih-Cheng Chen

Vision-based 3D sensing provides an effective approach for constructing human-centric digital twins in applications such as ergonomic assessment and human–robot collaboration. However, reconstructing reliable dynamic human digital twins from visual sensor observations remains challenging due to large articulated deformation, which often causes surface discontinuities in existing 3D Gaussian Splatting (3DGS)-based human avatars. Although current methods achieve high image-level rendering quality, their Gaussian primitives are not explicitly adapted to pose-induced surface deformation, resulting in cracks and holes during motion. In this work, we propose ElasticGS, a geometry-consistent human digital twin reconstruction framework for vision-based 3D sensing. ElasticGS introduces a training-free Pose-Aware Dynamic Covariance Adaptation strategy, which dynamically adjusts the rotation and anisotropic scale of Gaussian primitives according to local human surface deformation. Specifically, localized Principal Component Analysis (PCA) is performed on deformed SMPL-X mesh neighborhoods to estimate surface stretching directions and magnitudes, enabling Gaussian representations to maintain continuous surface coverage under complex poses. Furthermore, we propose the Avatar Surface Integrity Score (ASIS), a reference-free metric that evaluates structural completeness and surface compactness of reconstructed human avatars without requiring pixel-aligned ground truth. Experiments on the X-Humans, AvatarReX, and People-Snapshot datasets demonstrate that ElasticGS significantly improves surface integrity while maintaining competitive image-level sensing consistency. Compared with ExAvatar, ElasticGS maintains satisfactory rendering quality while improving ASIS from 0.741 to 0.961 on X-Humans, from 0.871 to 0.949 on AvatarReX, and from 0.874 to 0.952 on People-Snapshot. These results demonstrate the effectiveness of pose-aware covariance adaptation for robust human digital twin reconstruction in vision-based 3D sensing systems.

Sensors doi: 10.3390/s26217013

Authors: Ernestas Nacius Vladislovas Čižas Kasparas Stanaitis Ignas Grigelionis Linas Minkevičius

Terahertz and sub-terahertz links offer wide bandwidths for future wireless communication, but practical deployment is limited by high propagation loss, atmospheric absorption, and sensitivity to blockage in non-line-of-sight environments. Here, we demonstrate ultrathin, passive full-space beam-shaping elements based on complementary split-ring resonators laser-cut into stainless steel foil. By encoding the target phase profile in the spatial arrangement of identical complementary split-ring resonator meta-atoms, which act as a binary amplitude-type zone plate, we shape the wavefront simultaneously in transmission and reflection and form foci on both sides of a single element. We show that off-axis routing (e.g., toward 45∘) is achieved via pattern arrangement without re-optimizing the complementary split-ring resonator meta-atoms for each geometry, providing a straightforward route to dual-channel routing. As a proof of concept, we transform a collimated incident field into Gaussian and Bessel-type beams and compare their propagation characteristics with more classical Fresnel-lens-like reflectors. These results suggest a potential low-profile approach to passive sub-THz beam control, which may support alignment-tolerant operation and could, in future work, be extended toward non-line-of-sight channel conditioning and compact multi-focus beam shaping.

Sensors doi: 10.3390/s26217012

Authors: Nadezhda S. Komova Kseniya V. Serebrennikova Anatoly V. Zherdev Boris B. Dzantiev

Simultaneous determination of salivary cortisol and testosterone levels could provide complementary information on the physiological status of the human body. However multiplex immunoassays are challenging because the two analytes occur at different concentration levels and the corresponding immunochemical reagents can interfere with one another. Here, we developed a duplex lateral flow immunoassay (LFIA) for simultaneous colorimetric determination of cortisol and testosterone. The assay was constructed on a single nitrocellulose membrane containing two spatially separated competitive immunoreaction zones. A distinctive feature of the immunoassay design was a selective interception zone that was located between the two test lines and functioned as an immunochemical sink. The optimized assay provided working ranges of 0.38–5.41 ng/mL for cortisol and 27–845 pg/mL for testosterone, with calculated detection limits of 0.18 ng/mL and 10.1 pg/mL, respectively. The analysis took 20 min and enabled the detection of both signals using a single test strip and colorimetric signal processing. Analysis of spiked artificial saliva samples yielded recoveries of 96.4–105.7%. The results demonstrate that spatial organization of immunochemical reactions may represent an alternative strategy for multiplexing LFIAs in cases where spectrally distinct labels are undesirable or when the analytical requirements of individual assays differ significantly.

Sensors doi: 10.3390/s26217011

Authors: Thao Baszenski Jan-Christoph Möller Karl-Heinz Kratz Georg Jacobs Tobias Gemmeke Anuj Khare Benjamin Lehmann Mattheüs Lucassen

Hydrodynamic plain bearings are widely used in heavy-duty applications, e.g., wind turbine drivetrains or ship propulsion systems. However, abnormal events such as mechanical overload causing overheating or particle contamination can lead to rapid bearing damage and the failure of the entire drivetrain. Most existing condition monitoring systems (CMS) cannot detect particle contamination and mechanical overloading in a timely manner or are associated with high costs and installation effort that limit their practical applicability. In this paper, temperature field measurement (TFM) is investigated as a promising approach to address this gap via the development of a sensor-integrating plain bearing (SiPB) prototype. The approach is characterized by its low cost and low energy demand, with the potential for autarkic operation through the thermoelectric harvesting of the frictional waste heat within the bearing. In both overheating tests, characteristic signal discontinuities in the temperature field occurred at least 10 min before the friction torque limit was reached, which served as the criterion for terminating the experiment. Particle contamination resulted in a pronounced increase in the scattering in the temperature field over the entire duration of the experiment compared to regular operation. The interquartile range (IQR) of the mean temperature was identified as a promising characteristic; in the single-contamination test, the IQR was increased by a factor of 2.6 compared with regular operation. These characteristics were derived descriptively from one test per abnormal event; repeated tests are required to validate them. The energy demand of the SiPB prototype is determined as a function of the sampling rate (1.3 mW at a sampling rate of 0.2 Hz). The temperature differences required for cold start and the supply mode of the DC-DC converter necessary for supplying the required operating voltage of the SiPB prototype via thermoelectric energy harvesting are determined (approximately 2 and 5 K, respectively). Based on simulation data, this corresponds to the magnitude of temperature differences expected in the plain bearing during normal operation, meaning that the autarkic operation of the SiPB can be enabled in the future.

Sensors doi: 10.3390/s26217009

Authors: Qi Lu Zhilin Li Dongxiao Pang

Missing values, outliers, out-of-order timestamps, unit inconsistencies and baseline drift degrade online monitoring data on drilling-fluid properties in deep wells. We validate an evaluation–marking–remediation–validation workflow on a multi-source database of about 500,000 business records and on 150,000 instrument and replay records that are registered separately by source and are not summed, from five wells. Six metrics describe data quality: missing rate, out-of-order rate, outlier rate, unit consistency, baseline drift and noise. Missing values are treated in tiers: short gaps are interpolated, medium gaps are interpolated as last-observation-carried-forward with a mask, and long gaps invalidate the window. Offline filling uses forward filling, linear interpolation or random forest prediction. MAD/Hampel detection with operating-condition matching separates sensor artefacts, operational transients and genuine anomalies. Signal enhancement combines robust denoising, multi-scale fusion and a trend-confidence threshold. On the same replay dataset, the missing rate fell from 6.8% to 0.7%, the outlier rate decreased from 3.6% to 0.8%, the baseline-drift index fell from 1.47 to 0.61, the noise level decreased by 40.8%, and the usable-record ratio rose from 77.0% to 98.0%. On a fixed judgement set of 55 genuine events and 164 negative exposure intervals, the operator that retains operations and multi-metric co-movement recovered 55 of 55 genuine events at a negative-interval false alarm rate of 25.61%, against 0% recall for the 3-sigma rule and 54.55% recall for Hampel alone; this comparison covers simplified cleaning operators and is not equivalent to the complete remediation chain. Warning models trained on the remediated data reached an F1 of 93.0–94.0% and lead times of 1.8–2.8 h on the replay test dataset. The raw data replay and the field application are two different data sources and are not compared as an effect of remediation.

Sensors doi: 10.3390/s26217010

Authors: Vessela Krasteva Todor Stoyanov Irena Jekova

Neurodegenerative diseases (NDDs), including Parkinson’s disease (PARK), Huntington’s disease (HUNT), and amyotrophic lateral sclerosis (ALS), are associated with characteristic gait alterations. This study proposes an image-based deep-learning approach for NDD gait classification and its implementation on an edge-computing platform. The main contributions include: (1) mapping synchronized left- and right-foot ground reaction force (GRF) signals from individual gait cycles into a common 2D bilateral force space to generate three complementary representations—Trajectory Map (bilateral force trajectories), Derivative Map (dynamic transitions), and Occupancy Heatmap (loading-state distributions); (2) adapting ImageNet-pretrained MobileNetV2 to four-class gait classification (CONTROL, ALS, HUNT, and PARK) through classification-layer retraining and feature-extractor fine-tuning; (3) comparing compact (α = 0.35; ~0.4 M parameters) and higher-capacity (α = 1.4; ~4.4 M parameters) MobileNetV2 configurations; and (4) evaluating the deployment of the resulting models on a Raspberry Pi 4 as a resource-constrained edge-computing platform. The models were evaluated on a test subset of 2259 gait cycles from the Gait in Neurodegenerative Disease Database, comprising CONTROL (612), ALS (367), HUNT (727), and PARK (553). Fine-tuning increased overall four-class accuracy from 89.64–92.61% to 95.09–96.41% across the GRF representations and model configurations. Paired comparisons showed improvements of 3.81–6.29 percentage points across all six matched configurations. Class-specific one-vs-rest accuracy was consistently high after fine-tuning: CONTROL (97.74–98.63%), ALS (98.41–98.94%), HUNT (96.02–97.21%), and PARK (97.30–98.23%). The highest four-class accuracy (96.41%; 95% CI: 95.45–97.29%; Occupancy Heatmap, α = 1.4) was preserved for the unquantized TFLite model evaluated on the Raspberry Pi 4, while dynamic-range quantization reduced accuracy by 0.22 percentage points to 96.19%. Compact fine-tuned MobileNetV2 models required approximately 42–46 ms of inference latency and 89–97 mJ of estimated incremental energy per inference, compared with 324–374 ms and 680–785 mJ, respectively, for the higher-capacity models. Dynamic-range quantization reduced model storage by approximately 62% (α = 0.35) and 72% (α = 1.4). Overall, bilateral GRF representations combined with MobileNetV2 transfer learning enable accurate multiclass NDD gait classification, while compact models provide a favorable performance–efficiency trade-off for resource-constrained edge implementation. However, subject-independent validation remains necessary because the reported accuracy was obtained using gait-cycle-level data partitioning.

Sensors doi: 10.3390/s26217008

Authors: Dong Ni Yu Fang Peng Zhang Xinguo Zhu

Air surveillance radars typically operate in two distinct modes—omnidirectional surveillance and sector-specific surveillance—which are inherently incompatible in conventional designs. To overcome this limitation, this paper proposes a novel flexible detection technique for electro-mechanical scanning radars that integrates omnidirectional surveillance with sector-specific energy focusing. The operational constraints and capacity boundaries of the proposed hybrid mode are rigorously analyzed and quantified through a mathematical framework. Furthermore, UAV-based calibration flight tests are conducted to validate the feasibility of the proposed technique. Experimental results demonstrate that the hybrid mode achieves a significant detection range extension in the energy-focused sector while maintaining continuous omnidirectional coverage in other sectors. This approach therefore offers a practical and effective novel operational mode, which can significantly enhance the battlefield adaptability and detection performance of electro-mechanical scanning radars.

Sensors doi: 10.3390/s26217007

Authors: Irfan Ullah Farhan Ullah Khalil Khan Rehan Ullah Khan JaKeoung Koo

Proximal, ground-based infrared thermography of substation equipment grades a temperature rise computed under a single, camera-default emissivity and reflected-temperature setting applied to every pixel, regardless of the material it shows. This assumption is inexpensive but not harmless: two components at the same physical temperature but emissivities of 0.3 and 0.9 are read 24.0 K apart, enough to shift a component between adjacent bands of the ANSI/NETA MTS maintenance-priority scale. The measurement equation responsible sits on the same physical basis as land surface temperature retrieval from satellite and airborne thermal sensors, where emissivity is likewise the term most often assumed rather than measured. This paper corrects and validates that assumption at corpus scale by fusing the thermal frame with what a modern radiometric camera already records alongside it, a co-registered visible-light photograph, on a public 895-frame corpus of substation thermograms acquired under a single fixed set of acquisition parameters. The visible channel is used to rebuild the radiometric measurement equation, recover a per-frame reflected temperature from sky segmentation, and estimate in-situ emissivity at 106 metal-to-insulator material boundaries, with the resulting uncertainty propagated by a per-frame Monte Carlo to the severity band assigned to 765 graded frames. Three reference-free tests, a within-insulator negative control, a joint emissivity-reflectance fit, and a cross-view repeatability analysis, validate the estimated emissivity against physical constraints rather than a laboratory thermometer. Under the conventional absolute-temperature criterion, 48.4% of frames cannot be assigned a severity band with 80% confidence even once air temperature is obtained from meteorological reanalysis rather than a fixed default; grading instead against a photometrically similar in-frame reference surface reduces this to 5.9% (12.2% once the reference surface’s own sampling noise is additionally propagated through the same Monte Carlo). A value-of-information ranking identifies the single additional measurement, by material, that would most reduce this ambiguity. These results argue for relative grading criteria over a more accurate default emissivity table, an approach tied to the visible-thermal pairing common to modern radiometric cameras rather than to a specific camera model or acquisition site.

Sensors doi: 10.3390/s26217006

Authors: YoungKyun Oh SeongKeun Jang BeomKyu Suh Ki-Il Kim

Mobile sinks have been widely studied to mitigate the energy-hole problem and improve data collection in wireless sensor networks. Although existing surveys have primarily examined sink mobility, routing, and energy efficiency, they provide limited analysis of the interdependent coordination and optimization problems arising from multiple mobile sinks. To address this gap, this survey presents a coordination-centered review of multiple-mobile-sink-assisted data collection, organizing existing studies into five research dimensions within three complementary analytical levels: decision problems, decision coupling, and solution paradigms. These dimensions encompass task assignment and network partitioning, coordinated trajectory planning, multi-sink routing and data collection, cross-layer coupling and joint optimization, and intelligent multi-sink coordination. Beyond conventional classification, we analyze decision dependencies, distinguish sequentially coupled approaches from explicit joint optimization, and compare representative solution methodologies and their operational trade-offs. We further examine reinforcement learning and multi-agent reinforcement learning alongside alternative learning paradigms, distinguishing learning-based adaptation, federated parameter sharing, and explicit cooperation among multiple physical mobile sinks. We also identify open challenges involving learning scalability and heterogeneous multi-sink coordination, as well as security and practical deployment constraints.

Sensors doi: 10.3390/s26217003

Authors: George Koufoudakis

In Internet of Things (IoT) applications, measurements may be requested at specified points and assigned to worker nodes whose locations and reported battery states differ. For such workloads, GeoBattery was implemented as an optional Kubernetes Scheduling Framework plugin. Target coordinates are read from pod annotations, while fresh ordinal battery reports are obtained from node conditions. Eligible nodes are then ranked by a configurable combination of Euclidean distance and reported battery state, following the standard feasibility checks. The plugin was evaluated against generated workloads in a Go simulator without a running cluster. Proximity, battery, and combined policies were compared over 1000 timesteps per run. A default profile was examined separately, and sensitivity was studied by changing one factor at a time. Under combined scoring, intermediate distance and completion outcomes were obtained relative to the policies using either input alone, with small differences in modeled lifetime during observation. These findings are confined to the generated workloads and stated battery assumptions. Physical energy use, battery endurance, and sensing quality remain to be measured on devices, and scheduler behavior requires evaluation in a live cluster.

Sensors doi: 10.3390/s26217005

Authors: Rongbo Zhou Jiao Tian Liejun Wang

Federated human activity recognition (HAR) with wearable sensors is challenged by label-skewed client data, missing local activities, and scarce mini-batch negatives. We propose FedREPA, a relation-guided framework that couples missing-class-aware relation consistency with prototype contrastive learning. Clients derive an inter-class relation prior from shared global class prototypes. A masked relation-consistency loss matches supported local relation rows to this prior. The same prior reweights negative classes in the prototype contrastive loss, giving greater emphasis to nearby prototypes when trust is held fixed. Global prototypes provide cross-client positive and negative anchors, including negatives for activities absent locally. On UCI HAR, PAMAP2, CIFAR-10, and CIFAR-100 under simulated Dirichlet label skew, FedREPA records the highest observed mean global test accuracy in the comparison with the ten evaluated baselines at each tested Dirichlet level, using the last five evaluation points per run. FedTGP, FedDPA, and FedAli-PHAR were not evaluated, and no performance advantage over them is claimed. Its paired accuracy margin over the restricted FedGPS-Core baseline averages about one percentage point. Worst-client accuracy is also highest among the methods shown on UCI HAR and CIFAR-10, while selected-class recall on CIFAR-10 shows mixed results. These findings are descriptive comparisons under the evaluated Dirichlet label-skew protocol. Performance under subject-wise federated partitions, device-wise heterogeneity, and sensor-placement variation remains unvalidated, as does clinical applicability.

Sensors doi: 10.3390/s26217004

Authors: Zixuan Liu Wei Xiong

Microphone-array direction-of-arrival (DOA) estimators must cope with changes in array layout, sensor availability, and the number of active sources. We present GCPR-DOA, a geometry-conditioned pair representation that encodes each cross-spectral pair together with its microphone baseline and frequency before masked aggregation, so that one network processes arrays with 4 to 32 microphones. Two decoders use this representation: a set decoder supervised through analytic positive-semidefinite covariance reconstruction, and a spectral-moment decoder that forms a geometry-weighted steered coherence spectrum and counts sources from energy-aware features. On the LOCATA corpus, paired three-seed ablations show that geometry conditioning lowers the optimal subpattern assignment (OSPA) error by 1.35 degrees and the fixed-output angular error by 9.7 degrees. Trained on the same queries as an adapted SRP-DNN and evaluated on 46,457 common queries, the spectral-moment decoder reaches median single-source errors of 5.5–5.8 degrees, compared with 8.8–9.0 degrees, and higher active-query F1 (0.51–0.54 versus 0.47–0.49); on the three arrays other than DICIT, the two methods are comparable (5.7–5.8 versus 6.3 degrees). On the original 33,497-frame evaluation, localized F1 at ten degrees rises from 0.137 to 0.568–0.597, above all classical references.

Sensors doi: 10.3390/s26217002

Authors: Yunsu Bae Byung-Jun Jang

IEEE 802.15.4z High-Rate Pulse (HRP) Ultra-Wideband (UWB) wireless technology supports centimeter-level ranging and is widely utilized for indoor positioning. Conventional single-anchor systems typically combine two-way ranging (TWR) for distance measurement with phase difference of arrival (PDoA) processing for direction estimation. However, PDoA accuracy depends strongly on the array aperture, angle of arrival, and received signal quality, which inherently limits the effective positioning range of the system. To address this limitation, this paper proposes a single-anchor positioning system based on an N-element circular array of directional antennas, in which a single UWB transceiver sequentially switches among the antenna elements. The anchor determines the distance via double-sided two-way ranging (DS-TWR) and estimates the direction based on the relative power between channels, obtained by measuring the first-path power of each antenna channel. Experimental evaluations conducted in a 12.8 m × 22.8 m indoor hall resulted in median and 90th-percentile direction errors of 1.93° and 2.66°, respectively, and median and 90th-percentile positioning errors of 0.24 m and 0.36 m, respectively.

Sensors doi: 10.3390/s26217000

Authors: Erika De Bardi Giambattista Gruosso Giovanna Dondossola Roberta Terruggia

Modern Electric Vehicles (EVs) and their charging points (CPs) function as Internet of Things (IoT) nodes embedded in the distribution grid: they continuously sense local electrical conditions, exchange real-time data with the Charging Station Management System (CSMS) via the Open Charge Point Protocol (OCPP), and adjust their power consumption in response to remote commands. This dual role as both a controllable load and a networked sensor makes EV charging infrastructure a high-value target for cyber–physical attacks. This paper investigates the impact of open-loop multi-point Dynamic Load Altering Attacks (D-LAAs) carried out by exploiting this IoT interface on a modified IEEE 13-node distribution feeder hosting twenty CPs at one node. A high-fidelity cyber–physical co-simulation testbed is proposed, coupling a real-time Hardware-in-the-Loop (HIL) power grid model implemented in MATLAB/Simulink and deployed on an OPAL-RT OP4512 platform via RT-LAB, with an emulated ICT network running in Mininet on Ubuntu 22.04. OCPP 1.6 over WebSockets governs CP-CSMS communication, while a TCP/IP interface links the two simulation domains. Two attack scenarios are studied: a coordinated step attack, in which all CP setpoints are first driven to zero and then forced to maximum power (120 kW), and a periodic modulation attack, in which setpoints are rapidly alternated between zero and maximum. Voltages are evaluated at three nodes with varying topological distances from the attack point. In both scenarios, voltage drops reach up to 3.74%. This violates standard voltage-quality margins, demonstrating that corrupting the IoT sensing-and-actuation layer of EV fleets is sufficient to threaten distribution grid stability and underscoring the need for dedicated anomaly-detection and mitigation strategies.

Sensors doi: 10.3390/s26217001

Authors: Jongyun Kim Seongkeol Kim

Laser-speckle tracking enables non-contact vibration measurement, but modal identification during rotation remains challenging. This study examines whether resonance-like frequency response functions and high coherence provide sufficient evidence of modal identification. A steel shaft was tested at 0, 100, and 200 rpm using impact excitation at four positions. The assessment combined stationary cross-validation against an accelerometer, matched pre-impact comparisons, and measurement of optical degradation. At 0 rpm, optical frequencies differed from contact-reference values by 1.06% at P2 and 2.60% at P3. The two rotating speeds and four excitation positions produced eight measurement conditions. Two yielded no candidate within the decision band. Six conditions yielded optical candidates. In three of these conditions, coherence at the selected frequency was higher before impact than after impact. Removing a linear phase trend eliminated one additional candidate; however, this outcome varied in leave-one-out tests. Two tracking methods produced candidates 12.21 Hz apart at 200 rpm, but the cause remained unresolved. Rotation increased the optical noise floor by 35–46 dB and reduced speckle correlation from 0.999 to 0.708. Under these conditions, high coherence alone did not validate modal identification. Matched pre-impact comparisons provide an additional check, but passing this check does not confirm a structural mode.

Sensors doi: 10.3390/s26216999

Authors: Dani Marchuet Javi Palanca Vicent Botti

This paper proposes a noise-aware, iterative methodology for enriching an initial mobility clustering through the sequential addition of explanatory variables. The framework separates one-off density-based cluster discovery from scalable characterisation over a fixed analytical population. Candidate variables are retained only when they improve the predefined distributional criteria and preserve continuity with the latest accepted partition. Evaluation on a large sensing-based mobility dataset from Valencia shows that the proposed architecture supports scalable and auditable explanatory enrichment while keeping the observation set unchanged and explicitly separating distributional admissibility from inter-partition continuity.

Sensors doi: 10.3390/s26216996

Authors: Daniela Oliveira Lucas Veras Hélder Fonseca

Accelerometers are widely used to estimate energy expenditure (EE) and classify physical activity intensity (PAI), although conventional approaches may not capture the complexity of raw acceleration signals. This study evaluated machine learning (ML) models for estimating EE and classifying PAI during daily and locomotor tasks. Forty-nine adults (28.4 ± 8.5 years) completed seven laboratory activities while wearing ActiGraph GT9X Link accelerometers on the right hip and dominant and non-dominant wrists. EE was measured by indirect calorimetry. Time- and frequency-domain accelerometry features from non-overlapping 1-s windows were combined with age, sex, body mass index (BMI), and body mass. K-nearest neighbors (KNN) and random forest (RF) models were developed for each placement using an 80/20 training–test split. RF achieved lower overall EE estimation errors and higher overall PAI classification agreement than KNN across placements. The lowest EE errors were obtained from RF models at the non-dominant wrist (RMSE = 0.67 kcal·min−1; MAE = 0.39 kcal·min−1) and dominant wrist (RMSE = 0.68 kcal·min−1; MAE = 0.40 kcal·min−1). RF classification agreement reached 95.81% (Kappa = 0.943) at the non-dominant wrist and 94.97% (Kappa = 0.931) at the hip. In this laboratory protocol, the non-dominant wrist RF model showed a favorable combination of EE estimation error and PAI classification agreement.

Sensors doi: 10.3390/s26216991

Authors: Liang Wu Shengyong Xu

Terrestrial bioinspired robots emulate the locomotion strategies of land animals and have potential applications in disaster response, planetary exploration, and inspection of complex terrain. This review examines the biological basis of six representative locomotion modes: multi-legged walking and running, elastic jumping, ground-coupled gliding, interface-mediated climbing, peristaltic crawling, and adaptive morphogenesis. It introduces dimensionless descriptors of their kinematics, with explicit reference scales for periodic motion and one-off events. We use morphology-function correspondence to compare representative multi-legged, quadrupedal, jumping, gliding, climbing/peristaltic, and multimodal robots with respect to mechanisms, actuation, sensing, control, quantitative performance, and application scenarios. The cited demonstrations range from laboratory mechanisms to selected quadrupedal field applications; their heterogeneous validation records do not support class-wide numerical readiness ratings. Incomplete reporting of sensing and endurance also limits comparison. The review identifies task-specific co-design needs, with sensing architectures matched to the contact and transition variables that must be regulated in each locomotion mode.

Sensors doi: 10.3390/s26216998

Authors: Dang Bui Hai Phat Nguyen Khac Trung Van Duc Quang Vu Van Thang Vu Toan

This paper proposes a physics-informed identification framework for the spatial and dynamic responses of dual-element pyroelectric infrared (PIR) sensors using variable circular-aperture excitation. The dynamic half is carried through to experimental identification, while the spatial half is carried only to the reference-pose location, so the joint estimation is proposed rather than demonstrated. The spatial response is modeled as a bipolar differential kernel composed of a pose-dependent optical mapping, an optical envelope with a spatial scale factor, two smoothed rectangular sensing-element footprints, and a small antisymmetric residual correction. Sinusoidal modulation of the aperture radius leads to a complex first-harmonic frequency-domain model relating the measured response to aperture radius, modulation amplitude, pose setting, and excitation frequency. Rather than being assumed, the dynamic structure is identified from a 40-point frequency sweep spanning 0.0999 to 4.0453 Hz. A compact two-time-constant differentiating model leaves a root-mean-square log-magnitude error of 15.1%, with a maximum log-magnitude deviation of 48.4% (62% in relative amplitude). A structure with two zeros at the origin and four poles reproduces the same measurement to 0.71% root-mean-square. Its poles are an underdamped complex pair at 0.2126 Hz with damping ratio 0.899 and real poles at 0.5023 Hz and 2.9672 Hz. The additional zero and pole are tentatively attributed to the AC-coupled conditioning stage. Evaluated at this identified operating point, the scaled Jacobian of the dynamic block is full column rank with cond(Js) = 75.8, and Cramér–Rao bounds agree with residual-bootstrap intervals to within ten per cent on every parameter. A greedy D-optimal selection retains 61.6% of the information of the full sweep from ten frequencies, whereas the originally proposed excitation grid attains only 13.8% because it does not sample the band containing the identified pole pair. Restoring the excitation phase reference increases the Fisher determinant by a factor of 130. A structured estimation procedure is also formulated, in which the residual coefficients are updated through a regularized linear least-squares subproblem within a staged optimization exploiting the disparate conditioning of the spatial and dynamic blocks. It has not yet been executed on a joint aperture–pose–frequency data set.

Sensors doi: 10.3390/s26216997

Authors: Haiquan Cao Meiqi He Jingyuan Wang Xiaowu Chen Zhen Liu

Design for testability configures test points and plans tests during product design to support fault detection and isolation in service. Existing test strategy generation (TSG) methods mainly address offline diagnosis. As demand for uninterrupted operation grows, downtime and economic losses become more significant, shifting maintenance toward online diagnosis. However, offline methods optimize complete diagnostic processes without directly accounting for isolation gains within a limited time. Their direct application can therefore yield a low fault isolation rate (FIR). To address this problem, this paper proposes an online TSG method based on delayed-feedback value learning (DFVL). First, an optimization problem is formulated using a multi-signal flow graph and a finite time constraint. Second, a DFVL framework is developed using value recursion and feasible subtree returns. It comprises Bellman-based features, balanced feasible completions, delayed updates with unweighted normalized residuals, historical-tree retention, model-derived MAP candidates, and bounded local search. Across 90 synthetic models and 12 budgets, mean FIR was 83.53%, compared with 77.96% for Mix1 and 73.24% for DIG. At normalized Time 36, device-model FIR was 75.90%, versus 53.15% for Mix1 and 61.86% for DIG. FIR denotes a model-based correct terminal MAP decision probability under single-fault, deterministic-response, and fixed-duration assumptions.

Sensors doi: 10.3390/s26216995

Authors: Chunguang Wang Wei Sun Jianlong Zheng Meijing Ge Yuming Qi

Streaming feed-forward 3D reconstruction produces depth predictions at successive frames, but cross-view geometric consistency and local depth accuracy may vary. We present a training-free, output-level approach that combines Multi-Reference Geometric Risk assessment with global-referenced conservative sparse depth refinement, leaving the pretrained backbone, its predicted poses, and its streaming state unchanged. Risk derived from predicted depth, intrinsics, and poses achieves a median Spearman correlation of 0.635 with scale-aligned depth error across 11 public scenes and a correlation of 0.556 on the real-robot Sequence E. With matched 2% sparse depth anchors, local refinement reduces Sequence E held-out AbsRel by 40.69% relative to Sparse Global. Spatial refinement also improves mean held-out depth error in all 55 public scene–policy comparisons and all 200 predefined fixed-mask scene–policy–seed comparisons. Retrospective risk-based frame ranking provides diagnostic benefits under reference-scale normalization, but the observed allocation gains are protocol-dependent and do not establish consistent improvements under historical sparse-scale propagation or deployable online scheduling. In an independent Bonn Leica comparison using shared ground-truth camera poses, Spatial increases recall at a 5 cm threshold but has lower precision and F-score than Sparse Global. These findings establish the value of conservative sparse depth refinement and multi-reference risk as complementary geometric-quality evidence, while distinguishing depth accuracy from downstream 3D geometric performance.

Sensors doi: 10.3390/s26216994

Authors: Chao Pan Yong Qian Zehao Chen Xiaoxin Chen Jiuyi Zhao Gehao Sheng

When a lightguide rod extracts partial discharge (PD) light from gas-insulated switchgear (GIS), losses introduced by the rear-section fixture and aperture mismatch with the fiber entrance limit the usable optical output. Ray tracing was used to examine how a parabolic taper affects optical transmission and reception through a finite aperture. The selected design increased the mean radiant flux received at the fiber entrance to 7.26 times the baseline value. In integrating sphere tests under uniform illumination, the sample with a parabolic taper provided a higher detector output than the other three samples under all tested conditions. In needle–plate corona tests in a laboratory GIS model, the main optical responses corresponded in time to the ultrahigh-frequency (UHF) envelope and shared the same temporal features. The optical channel also recorded clusters of low-amplitude responses during the negative half-cycle that were not clearly resolved by the UHF channel; these responses were accompanied by activity in the auxiliary ultrasonic channel. These results identify rear-section transmission loss and finite-aperture reception as constraints on lightguide design and provide experimental evidence for coupling lightguide rods to transmission fibers in optical PD detection in GIS.

Sensors doi: 10.3390/s26216993

Authors: Nikolaos Pagonis Konstantinos Plomaritis Evangelos Kogias Nikolaos Foroglou Ioanna Chouvarda

Preoperative differentiation of glioma from meningioma on MRI is important for treatment planning but depends on expert interpretation. A 3D ResNet-34 with CBAM attention was trained on four-modality MRI from 2251 BraTS 2023 patients (1801 for four-fold cross-validation; 450 held out for testing) and achieved 96% accuracy, an AUC of 0.993, and an expected calibration error of 0.0165 on the test set. For layer-wise explanation, Grad-CAM and SmoothGrad were applied at the four residual stages, and a reconstruction decoder was added to obtain full-resolution, decoder-mediated attribution at the 4 × 4 × 4 bottleneck, where standard Grad-CAM is too coarse. Four quantitative metrics gave complementary results: the decoder-assisted attribution had the highest pointing-game accuracy (80.2%) and the lowest deletion AUC, whereas Layer 1 had the highest insertion AUC and Layer 2 the highest mean faithfulness delta, with no significant faithfulness difference among Layers 1, 2, and 4. Glioma was localized by most layers, while meningioma was localized mainly through the decoder-assisted pathway. A preliminary clinical review and a normalization analysis are also reported. The decoder offers an annotation-free way to obtain full-resolution attribution from deep layers of volumetric models.

Sensors doi: 10.3390/s26216992

Authors: Yibo Cao Jiahong Huang Zijun Yin

Vision-guided robotic manipulation requires control actions that reflect visual observations, robot state, and task goals. Action Chunking with Transformers (ACT) performs well on individual tasks, but shared multi-task policies must accommodate different action distributions and state-dependent control requirements. We propose an ACT-based selective residual action-chunking policy using global and wrist RGB images together with proprioceptive feedback. A task-conditioned ACT model is first trained on multi-task demonstrations. Its parameters are then frozen while bounded residual experts learn candidate action corrections. A gain predictor uses observation-conditioned decoder features, robot state, task information, and base-action statistics to select a correction or retain the base prediction. Gain supervision is derived from reductions in imitation error. The robot executes an action prefix and updates its prediction from new observations, incorporating selective correction into closed-loop manipulation. Across all ten LIBERO-Goal tasks, the method achieves a mean success rate of 91.2%, exceeding MTACT and BAKU by 11.4 and 4.2 percentage points. On three jointly learned real-world tasks, it achieves 78.7%, exceeding MTACT by 10.0 percentage points. These results support selective residual action generation for vision-guided multi-task robot control.

Sensors doi: 10.3390/s26216990

Authors: David Alexis Cano-Ibarra Wendy Flores-Carmona Karen Roa-Tort Diego A. Fabila-Bustos Macaria Hernández-Chávez J. Alejandro Aguirre-Anaya Josué D. Rivera-Fernández

Mexican Sign Language (MSL) plays an essential role in communication within the Deaf community, yet access to sign-language learning resources remains limited in some educational settings. This study presents a vision-based Human–Computer Interaction (HCI) system for MSL recognition and robotic reproduction, developed as a proof of concept toward inclusive educational applications. The system uses an RGB camera and MediaPipe Hands for hand-landmark extraction, a multilayer perceptron (MLP) for recognition of 29 static classes, and a TimeDistributed–Long Short-Term Memory (LSTM) model for 10 dynamic classes. Recognized signs are mapped to predefined motion sequences executed by an ATmega2560-controlled robotic arm. Under stratified random and sequential partitioning protocols, the static model achieved accuracies of 100.00% and 99.88%, respectively, while the dynamic model achieved 100.00% under both protocols. Additional comparisons with Support Vector Machine (SVM) and Random Forest classifiers, using three random seeds, demonstrated competitive performance and low variability within the original single-participant dataset. To assess generalization beyond the training participant, a preliminary signer-independent evaluation was conducted with three previously unseen participants, without model retraining or adaptation. Across 585 recognition attempts involving 39 classes, the system achieved an overall accuracy of 89.06%, providing initial evidence of recognition performance across different signers. Furthermore, 92 accepted recognition trials successfully triggered their corresponding robotic sequences, demonstrating functional integration between recognition and actuation. These findings support the technical feasibility of the proposed contactless HCI architecture as a potential foundation for interactive MSL learning tools. Nevertheless, broader signer-independent validation and user-centered evaluations are required to establish its accessibility, usability, and educational effectiveness.

Sensors doi: 10.3390/s26216989

Authors: Inchul Yang Hyoungsoo Kim Junhyung Lee

Lane selection and lane-change decisions by connected and autonomous vehicles (CAVs) require queue-length information at lane resolution, which corridor-level estimation methods do not provide. This paper presents the Lane-by-lane Queue Information Prediction System (LQS), which predicts per-lane queue length from upstream movement-traffic volume and signal-phase data by combining a platoon dispersion model with a shockwave-theory queue model on a rolling-horizon scheme. We describe its integration into a live C-ITS stack through an SAE J2735 message extension and its deployment at a testbed in Hwaseong, Republic of Korea, where processing latency was measured (313.6 ms average) but field accuracy was not. In an independent microsimulation testbed (Goyang, Republic of Korea), with parameters calibrated on signal cycles 6–15 and evaluated on held-out cycles 16–21, the mean cycle-maximum relative error on the four signal-controlled lanes is 24.5% (17.1–30.0%), against 33.2–52.8% for baseline methods; the right-turn lane, which discharges on red, is reported separately (51.5%). Because cycle-maximum and full-series errors rank lanes differently, we also report on RMSE normalized by mean queue, bias, and correlation, which reconcile the two metrics. An ablation shows that the two empirically tuned parameters are coupled.

Sensors doi: 10.3390/s26216988

Authors: Jiabao Xie Kuai Yu Tianfeng Hu Qingrui Yan Jiadong Zhang Wenguo Zhu Yongchun Zhong Jianhui Yu Yihua Qian Yaohong Zhao Huadan Zheng Tuan Guo

Dielectric immersion provides a direct thermal boundary for lithium-ion battery abuse tests, while passive fiber Bragg grating (FBG) sensing enables surface-temperature monitoring within the liquid. We monitored fully charged NCR18650BD cells during external short circuits under natural air and static transformer-oil immersion through two 1 km single-mode fiber leads. A seven-point temperature calibration extended to 132.3 °C yielded a sensitivity of 8.76 pm °C−1 (R2 = 0.99955). In the original tests, peak surface temperatures were 132.3 °C in air and 51.1 °C under immersion, with post-peak half-decay times of 579 and 161 s, respectively. Additional tests reproduced the lower temperature excursion and shorter post-peak half-decay under immersion, although the recorded short-circuit currents differed between cooling conditions. The results characterize these thermal responses and demonstrate the feasibility of electrically passive, long-reach FBG monitoring during immersed battery-abuse experiments.

Sensors doi: 10.3390/s26216986

Authors: Oskar Seifert Marcel Geis Niklas Alexander Koehler Manuel Stenger Volker Groß Martin Kramer Keywan Sohrabi

Time-of-flight (ToF) depth imaging offers contactless observation of respiratory surface motion in veterinary patients. We investigated the technical feasibility of extracting canine respiratory waveforms and explored agreement between recorded camera-derived respiratory rates and concurrent manual counts. Depth images were acquired at a nominal 10 frames/s, and a fixed region of interest (ROI) supplied a mean depth signal. Clinical rates were obtained using Fourier spectral analysis with interpolation. Four dogs contributed eight sessions and 40 paired one-minute observations. Mean bias was +2.33 breaths/min (95% confidence interval −3.60 to +8.25), mean absolute error was 3.93 breaths/min, and exploratory repeated-measures limits of agreement were −10.40 to +15.05 breaths/min. One dog accounted for 69.4% of the absolute error and all observations above 25 breaths/min. Supplementary analyses compared spectral and peak-based estimates on identical inputs. An additional analysis of 27 ROI configurations in each of two recordings from one dog found maximum minute estimate changes of 7.50 and 7.80 breaths/min relative to a fixed baseline ROI. These findings demonstrate waveform extraction and dependence on spatial and temporal processing. They do not establish reliable respiratory rate measurement across dogs or respiratory rate ranges. The small independent sample, heterogeneous errors, incomplete original FFT configuration records and unquantified reference timing constrain the clinical comparison.

Sensors doi: 10.3390/s26216987

Authors: Karina Awad-Pérez Maria Roldan Panicos A. Kyriacou

Optical brain monitoring techniques such as near-infrared spectroscopy (NIRS) and cerebral photoplethysmography (PPG) are promising non-invasive approaches for cerebral haemodynamic assessment; however, the measured signal can reflect contributions from both cerebral and extracerebral circulation. This study investigated the effect of extracerebral layers on forehead PPG signals using a multilayer head phantom with independent control of scalp and cerebral perfusion. Three perfusion conditions were performed: both layers perfused, scalp occlusion, and brain occlusion. PPG signals were acquired at 770, 810, and 880 nm using a custom optical probe with proximal (10 mm) and distal (30 mm) source–detector separations. Signal features were extracted and compared across conditions. Scalp occlusion produced a strong reduction in pulsatile signal, with complete signal collapse at the proximal detector. A small pulsatile signal (~5% of baseline amplitude) was detected at the distal detector, indicating that longer source–detector separations (30 mm) can detect signals originating from the brain phantom. Brain occlusion resulted in a partial amplitude reduction (~45%) with preserved waveform morphology across all wavelengths. While this demonstrates that cerebral perfusion changes are detectable in the presence of intact extracerebral circulation, it also highlights a potential masking effect as scalp circulation alone can produce a good-quality PPG signal despite complete cerebral hypoperfusion. These findings have clinical implications and highlight the need for mitigation strategies that account for extracerebral contributions in brain optical monitoring techniques.

Sensors doi: 10.3390/s26216985

Authors: Wen Yang Yelin Qi Hao Wen Yutong Pan Ting Liang

Unmanned aerial vehicles (UAVs) and unmanned ground vehicles (UGVs) equipped with chemical sensors hold great promise for hazardous incident reconnaissance, yet practical deployment is hindered by sensor survival upon impact, rotor downwash interference, and terrain-induced trade-offs between accuracy and efficiency. In this study, we focus on UAV-based deployment but also consider UGVs as complementary platforms for ground-level placement and long-term monitoring. To address these challenges for typical operational scenarios, we first implemented two customized engineering improvements: a carbon-fiber anti-downwash plate and a reinforced tether–release mechanism. Using a custom hexacopter with a 1 kg sensor prototype, we then systematically evaluated three deployment strategies. Direct free-fall from 5 m resulted in 0% survival due to housing fracture and circuit damage. The anti-downwash plate reduced local airflow by 70%, establishing a 6 m safe tether length. Tethered hovering achieved 100% survival, ~3 min/point mission time, and strong terrain adaptability, albeit with moderate dispersion. Low-altitude gentle placement delivered the highest accuracy and 100% survival but required 8 min/point and flat terrain. Based on these experimental results, we derive a mission-adaptive decision matrix for four representative chemical scenarios, recommending tethered hovering as the primary strategy for complex terrains, low-altitude placement for fixed long-term monitoring on flat surfaces, and direct drop only if cushioning is added and height is reduced to ≤3 m. We further discuss how the decision matrix can be extended to multi-sensor UAVs, UGVs, and UAV–UGV collaborative systems. Our findings offer a practical, evidence-based operational framework tailored to unmanned chemical reconnaissance.

Sensors doi: 10.3390/s26216984

Authors: Lulu Huang Mingyang Yang Zepeng Ni Dongmei Gong Kuai Zhou Peng Zhao Feng Xu

In response to the demand for efficient automated inspection of parts in the field of intelligent manufacturing, this paper proposes a design method for a high-precision visual measurement automated inspection platform oriented toward intelligent manufacturing scenarios. First, specifications for the layout of fixture reference points were established. An automatically rotating fixture was designed, and a robotic system was integrated with precision measurement equipment to build an integrated inspection and analysis platform. Robot collision simulation analysis and measurement registration analysis were completed, and key parameters in the inspection process were analyzed in light of the diverse inspection requirements of smart manufacturing. An automated measurement and analysis system that integrates automated testing, data processing, and analysis has been developed. Finally, validation tests were conducted. The results demonstrate that the automated inspection platform developed herein can efficiently implement automated product inspections with favorable measurement repeatability and high inspection efficiency, improving inspection performance for smart-manufacturing applications. The proposed method has good versatility and adaptability and can offer technical support for automated inspection in smart manufacturing applications.

Sensors doi: 10.3390/s26216983

Authors: Chengcheng Liu Wei Zhou Yong Ma Yifan Hu

Carrier screening and observation geometry determine how a local magnetoelectric response contributes to a potential-difference readout. This study develops a finite-length one-dimensional analytical model for a weak PN-junction piezoelectric-semiconductor fiber between magnetostrictive layers subjected to a localized magnetic-field increment about a prescribed bias. Cross-sectional force balance, two-carrier drift–diffusion, blocking end conditions, and interface continuity yield a four-region solution that separates the junction reference from the magnetic increment. In the representative weak-junction case, the end-to-end difference is about 1% of the peak local magnetically induced potential, whereas observation pairs at the field-region edges and near the junction give differences about 200 and 50 times larger than the end-to-end value. For equal screening and a step field, the response of a fixed interior pair peaks when the field-region edges reach the observation points. Uniformly doped and symmetric weak-junction controls have identical magnetic increments under equal screening despite their different reference fields. Layer thickness modifies both magnetic conversion and carrier screening. The resulting continuum-model relations identify how excitation width, screening length, and observation position jointly govern internal magnetic-field readout.

Sensors doi: 10.3390/s26216982

Authors: Wei Li Shuxun Li Bojiang Yin Jian Zhang Jianlong Zhu Zhijun Lei

In aerospace, industrial automation, and intelligent equipment, electromagnetically driven valve control systems often suffer from low control accuracy and stalling during switching and servo operations because of strong nonlinearity, parameter uncertainty, frequent operating condition variations, and the empirical nature of conventional PID tuning. To address these issues, this study proposes a PID self-tuning control method based on Deep Deterministic Policy Gradient (DDPG) to improve the dynamic response and robustness of valve control systems. An electromagnetically driven control valve system suitable for both on/off valve modes is constructed, and an interaction framework between the reinforcement learning agent and the Simulink simulation environment is established. The proposed reinforcement-learning control strategy establishes a mapping from the measured system states to the PID gains, enabling the gains to be adaptively adjusted according to the current operating state. To improve training stability and sample efficiency, experience replay and target network soft updates are introduced, and an Actor–Critic architecture is adopted for policy learning and value evaluation. The proposed strategy is further implemented in both the simulation model and MCU controller software, and its performance is compared with traditional PID and fuzzy PID methods. Results show that, in on/off valve mode, the method achieves timed opening, holding, and closing within a preset cycle, while in servo valve mode, it enables accurate position control over a 0~20 mm valve core stroke, with a positioning error of 0.04 mm and a response time below 50 ms. Compared with traditional methods, it shows clear advantages in accuracy, response speed, and adaptability.

Sensors doi: 10.3390/s26216981

Authors: Paulina Lewińska Kamil Maciuk Łukasz Więckowski Paweł Postek Sorin Nistor Łukasz Borowski

Thermal monitoring of complex topographic objects requires accurate integration of infrared data with three-dimensional spatial information. However, registering thermal imagery with point clouds remains challenging due to limited spatial resolution, low contrast, and the lack of stable reference features in thermal images. This study presents a methodology for generating a thermal point cloud using active thermal markers as control points. The developed markers are based on temperature-controlled Peltier modules, providing stable and detectable thermal signatures independent of the natural temperature distribution of the investigated surface. Laboratory and field experiments evaluate marker stability, detection capability, and applicability for integration with Terrestrial Laser Scanning (TLS). An automated procedure was developed to identify thermal markers and transfer temperature information from infrared images to point clouds using a Direct Linear Transformation approach. The proposed workflow was tested on a coal spoil tip, where a combined spatially referenced model was made. The results confirmed that active thermal markers enable reliable thermal image registration and allow temperature values to be assigned to three-dimensional terrain models. The proposed approach can support monitoring of mining areas, detection of potential zones of heat accumulation, and early identification of areas at risk of spontaneous combustion.

Sensors doi: 10.3390/s26216980

Authors: Md Kamrul Islam Md Mainul Islam Jannatul Bushra Orpa Joobeom Yun

Cooperative localization lets unmanned aerial vehicles (UAVs) fuse absolute and relative measurements, but the message exchange of distributed solvers is exposed to compromised UAVs. We study distributed weighted-least-squares (WLS) localization with the alternating direction method of multipliers (ADMM) under Byzantine deception, disruption, false data injection (FDI), and adaptive attacks. To mitigate these attacks, we propose a blockchainassisted ADMM in which a permissioned blockchain records one signed state per UAV and iteration, which prevents equivocation, and a relative-measurement check with a calibrated threshold screens the accepted states; a symmetric edge update makes the method independent of UAV labels. When honest neighbors accept each other, the method converges once the screening decisions stabilize and reject every compromised state, and at any fixed point its deviation from the accepted honest-subnetwork WLS solution is bounded in proportion to the screening threshold. In 1000-UAV simulations with 100–300 compromised UAVs, the proposed defense reduces the FDI error from 23.6–69.9m to 0.49–0.95m and stays within 0.03m of the ideal WLS estimate under equivocating attacks, whereas an adaptive attacker that attempts to evade the threshold causes errors of 3.4–8.5 m. On a six-node Geth Clique network, blockchain-integrated runs reproduce the simulations to within 2 μm, and with a raised block gas limit, an iteration with 1000 submitting UAVs takes about 3.8 s.

Sensors doi: 10.3390/s26216979

Authors: Răzvan Gabriel Boboc Ali Vatankhah Barenji

Over the past decade, virtual reality (VR) has evolved from a niche entertainment medium into a mature and versatile technological ecosystem supporting scientific research, education, healthcare, industrial training, human performance analysis, and interactive system design [...]

Sensors doi: 10.3390/s26216976

Authors: Qiang Han Josh Dafoe Bo Chen

Modern robotic systems increasingly rely on transformer-based AI models to process sensor input and support high-level decision making. However, due to limited GPU resources in each ego robot, it is difficult to run a large model locally. Outsourcing the model inference to cloud providers could address this issue, which however suffers from limitations including unreliable network connection, high latency, and high monetary cost. In this work, we have developed an effective decision making framework for robotic systems by leveraging a few state-of-the-art technologies, such as collaborative inference and model context protocol (MCP). Our framework seamlessly integrates collaborative inference between edge servers and mobile robots, supporting inference of large models by relying on local resources instead of remote cloud-based resources. It excels in a few feature designs. First, it is designed for mobile robot deployments where the available computing resources may change over time. This is achieved by introducing “anchors”, adding redundant nodes to hold the model partitions, and switching among model variants of different sizes. Second, it can allow heterogeneous robotic peers to seamlessly join the system through MCP, a standardized interface for modern AI applications. We have implemented a prototype of the proposed framework and experimentally validated its efficiency.

Sensors doi: 10.3390/s26216978

Authors: Michael Leitner Kyriaki Kostoglou Selina C. Wriessnegger

EEG-based stress detection is increasingly relevant for intelligent sensing and human–computer interaction, particularly for systems that aim to adapt to users’ changing cognitive and affective states. While conventional EEG-based stress detection approaches mainly rely on spectral band-power features, fractal-dimension measures and Hjorth parameters may provide additional information about underlying brain dynamics. However, their relative contribution, spatial distribution, and frequency-band dependence remain insufficiently understood. In this study, pre-recorded EEG data from a stress-inducing experiment were analyzed using statistical feature pre-selection and multiple machine-learning classifiers. The contribution of different EEG features was evaluated across participants, while scalp topographical maps were used to characterize the spatial distribution of informative features. The results identify Hjorth activity, frontal β asymmetry, and the γ/θ ratio as among the most frequently selected features in the ANOVA-based feature-selection analysis. Combined feature-space configurations including Hjorth parameters and fractal-dimension measures achieved higher classification performance than the conventional-feature reference configuration. These findings provide further insight into the spatial and frequency-specific characteristics of EEG features associated with the investigated stress-inducing cognitive task and demonstrate the value of combining conventional spectral features with Hjorth parameters and fractal-dimension measures for physiological-state sensing. The proposed approach may support the development of computationally efficient EEG-based monitoring of stress-related cognitive states for intelligent and adaptive human–computer interaction, including neuroadaptive and context-aware systems.

Sensors doi: 10.3390/s26216977

Authors: Cong Xu Chuang Hei Mingzhang Luo

In slow formations, dispersive quadrupole guided waves are used in acoustic logging while drilling (LWD), and a global coherence maximum that can switch between competing arrivals is picked by conventional slowness–time coherence (STC) processing. In this study, we train a lightweight convolutional neural network (CNN) on kinematic synthetic gathers, regress the inter-receiver moveout directly from quadrupole array gathers, and convert the estimated moveout to an apparent quadrupole-mode velocity. The conversion uses an empirically set effective spacing of 0.0725 m (nominal spacing of 0.100 m), so independent calibration is required for the absolute velocity scale. Five-fold cross-validation yields a median relative velocity error of 2.51% ± 0.16%, and the held-out test yields 2.11%, with R2 = 0.947. On the same held-out gathers, a lower median error (1.57%) is attained by STC, whereas fewer gross errors (relative error > 20%; 0.50% versus 3.88%) are produced by the CNN; sharp degradation of the CNN is shown by stress tests when secondary events are stronger (As/At ≥ 0.75) or the SNR is lower (≤5 dB) than that in training. On a measured slow-formation interval, numeric and image-inversion median apparent velocities differ by 13.9% before and 0.05% after 2.5–4.5 kHz band-pass filtering (pointwise r = 0.36). A control trained on the measured STC log reaches r = 0.53 on held-out records, but only r = 0.20 under depth-blocked cross-validation. On the same 16-core CPU, the 266 KB model is 3.4× faster than STC (≈21× with the CNN on a desktop GPU, which is not hardware-normalized). The results support kinematic moveout regression with potential edge suitability; absolute formation shear-wave velocity is not independently validated.

Sensors doi: 10.3390/s26216975

Authors: Xiao Luo Shan Liao Xue Gong Zhenyu Liu Xueni Zhao Chenghao Li Ling He Heng Yin

Perceptual evaluation remains the clinical standard for postoperative cleft-palate speech, but objective patient-level assessment of overall speech intelligibility remains limited. We developed a patient-level three-class classification model using frozen Chinese HuBERT representations, principal component analysis (PCA), and regularized multinomial logistic regression. The study included 123 patients with cleft palate, including 100 in the development cohort and 23 in the held-out cohort. Each patient completed 27 standardized articulation tasks. HuBERT embeddings from each task were summarized by temporal mean and standard deviation and concatenated in a fixed clinical order to form a patient-level representation. Model development used patient-disjoint nested cross-validation followed by locked held-out evaluation. The primary model achieved 90.00% accuracy and a macro-F1 of 0.8659 in development and 86.96% accuracy, a macro-F1 of 0.8227, 93.75% balanced accuracy, and a quadratic-weighted kappa of 0.9043 in the held-out cohort. All three held-out errors involved patients with Poor intelligibility who were misclassified as Reduced intelligibility, with no direct Normal-to-Poor errors. Comparisons across six classifiers and four pretrained encoders showed no significant paired differences after Holm correction. Task-replacement analysis revealed heterogeneous reliance on individual articulation tasks. These findings support the feasibility of objective patient-level speech intelligibility assessment from standardized cleft-palate speech recordings.

Sensors doi: 10.3390/s26216974

Authors: Jing Ma Aihua Zhang Yongqiang Ma Liang Zhu

Multi-step prediction of side-wall root weld width is essential for intelligent monitoring and closed-loop quality control in ultra-narrow gap arc welding. To address limited physical interpretability of pure data-driven models and insufficient control response time from single-step prediction, a Physics-Informed Informer (PI-Informer) multi-step prediction model constrained by welding thermal-balance mechanisms is proposed in this work. Thermal prior knowledge is embedded into training via a physically constrained loss function to compensate for physical deficiencies in pure data-driven learning. Local transient features are extracted from welding electrical signals by a multi-scale 1D-CNN and further integrated with global temporal dependencies of molten-pool heat accumulation through the Informer encoder. Unlike the standard Informer, decoder initialization is implemented with fused features of electrical signals and process parameters, enabling multi-step prediction at three typical prediction horizons of t, t + 3, and t + 15. From ablation experiments, the optimal weight β = 0.05 is identified for the best trade-off between accuracy and physical consistency, and multi-step errors are effectively mitigated by the 1D-CNN module. Under identical conditions, PI-Informer attains lower RMSE, MAE and MAPE than vanilla Informer, GRU and LSTM across evaluation metrics, and yields the lowest batch-averaged physical residual among all compared models. A favourable balance between accuracy and physical interpretability is achieved, and a feasible solution is provided for intelligent perception and proactive quality control in ultra-narrow gap welding.

Sensors doi: 10.3390/s26216973

Authors: Jingwen Gong Junxing Zhang Chunjuan Bo

Car logo detection is a challenging fine-grained perception task characterized by small object scales and high inter-class similarity. In such scenarios, discriminative cues mainly reside in subtle local textures rather than global structures. Although State Space Model (SSM)-based detectors, such as Mamba-YOLO with its ODMamba backbone, provide efficient long-range dependency modeling, they often struggle to maintain spatial alignment and semantic consistency during multi-scale feature fusion. This limitation reduces their effectiveness in fine-grained discrimination. To address these issues, this paper proposes LVSS-Net (Lightweight Visual State Space Network), a unified framework with three key components. First, Multi-Scale Semantic Alignment (MSAA) preserves spatial resolution while enhancing semantic consistency across feature pyramid levels through adaptive normalization and attention-based fusion. Second, Dynamic Content-Aware Reconstruction (DCR) introduces learnable content-adaptive offsets to model variable sampling positions during feature upsampling, mitigating scale-induced information loss and boundary blurring. Third, Global–Local Saliency Modulation (GLSM) combines parameter-free global saliency modeling with local texture enhancement via a lightweight LTE module to capture fine-grained discriminative patterns. Extensive experiments on the VLD-11 fine-grained car logo dataset and the TT100K traffic sign detection dataset demonstrate that LVSS-Net achieves 98.3% mAP@0.5 with only 1.5 M parameters. The proposed method reduces model complexity by 73.5% compared to the Mamba-YOLO baseline while maintaining competitive detection performance. These results demonstrate its effectiveness in resource-constrained fine-grained detection scenarios.

Sensors doi: 10.3390/s26216972

Authors: Ziao Xu Fulin Wang Qi Zhou Zhijian Cai Feifei Guo Zengzhe Xi Jie Dong

The acoustic impedance mismatch between piezoelectric materials and water loads remains a bottleneck that severely hampers the advancement of underwater acoustic transducers. In this work, we propose a novel design that utilizes magnesium (Mg) alloy as the matching layer for PZT-5H/Epoxy 1-3 composite underwater acoustic transducers. To assess the suitability of Mg alloy as a matching-layer material and provide basic data for transducer design, the microstructural dependence of the elastic and acoustic properties is systematically established for solution-treated, extruded, and forged GWZ932 (Mg-9Gd-3Y-2Zn-0.5Zr, wt.%) alloys. It is determined that the elastic coefficients and acoustic performance of the three Mg alloys are essentially independent of grain size, second phases, and texture, whereas the acoustic attenuation coefficient exhibited an increasing trend with grain size at 20 MHz. The three alloys exhibit nearly identical acoustic impedances of 11.01 ± 0.03 MRayl (solution treated), 10.94 ± 0.02 MRayl (extruded), and 11.02 ± 0.03 MRayl (forged). The performance of PZT-5H/Epoxy 1-3 composite transducers with four different matching-layer designs, including single-matching-layer designs (Mg alloy or Al alloy) and dual-matching-layer designs (Mg alloy + polymer or Al alloy + polymer), is simulated using a finite element method (FEM). The results demonstrate that the transducer with the Mg alloy + polymer dual-matching-layer design exhibits both optimal transmitting voltage response (TVR) and receiving voltage sensitivity (RVS) bandwidths, with −6 dB transmit and receive bandwidths of 188.7 kHz and 156.6 kHz, respectively. These simulation results indicate that the Mg alloy + polymer dual-matching-layer design is promising for use in broadband piezoelectric underwater acoustic transducers. This work presents a novel Mg-alloy-based matching-layer scheme, offering a new choice for the matching-layer design of underwater acoustic transducers.

Sensors doi: 10.3390/s26216971

Authors: Stefano Longo Stefano Pau Riccardo Battioli Nicholas Toninelli Giuseppe Coratella Fabio Esposito Luca Maria Sconfienza Emiliano Cè

Stretching acutely modifies muscle–tendon mechanical behavior. Whether these effects vary across muscle regions and stretching modalities remains unclear. This study investigated the acute effects of passive static stretching (PS) of different durations and dynamic stretching (DS) on the regional gastrocnemius medialis (GM) shear modulus. Twelve healthy adults (4 women; mean ± SD: age = 23.9 ± 2.4 years, body mass = 69.8 ± 10.6 kg, height = 1.72 ± 0.08 m) completed three randomized conditions: 5 × 30 s PS (PS30), 5 × 45 s PS (PS45), and 5 × 30 repetitions of DS. Shear-wave elastography assessed the shear modulus at mid-belly (GMmid) and towards the myotendinous junction (GMMTJ) at 0° and 80% of maximal dorsiflexion range of motion (ROMmax). ROMmax, passive resistive torque (PRT), myotendinous junction displacement (ΔMTJ), muscle architecture, and maximal voluntary contraction (MVC) were also evaluated. The GMmid modulus decreased at 80%ROMmax after all stretching protocols (p < 0.05), with no changes at 0°. At GMMTJ, the modulus decreased after PS45 and DS at both ankle positions (p < 0.05); no significant changes occurred after PS30. After all protocols, ROMmax increased (p < 0.05), while PRT and MVC decreased (p < 0.05), ΔMTJ increased post PS45 (p < 0.05), and muscle architecture remained unchanged. These findings indicate that stretching-induced changes in GM mechanical behavior can be region and muscle-length dependent, while the distal response may also be influenced by stretching duration and type.

Sensors doi: 10.3390/s26216970

Authors: Hrithik Majumdar Shibu Seyed-Ali Sadegh-Zadeh Delaram Jarchi

Millimetre-wave radar offers contactless exercise sensing, while wearable inertial measurement units (IMUs) can supply privileged training supervision without being required at inference. We study radar-only exercise recognition on the mRI dataset (20 healthy subjects; ten exercises) and distinguish descriptive historical architectures from matched RTPD-Net versus constrained fine-tuning (CFT) comparisons. Historical RTPD-Net, PARETO-RKD, and AURORA-RKD were trained using different budgets, seeds, or corruption settings and cannot establish a preferred method. For a fixed 40-segment test split, the original evaluation reported mean corrupted macro-F1 changing from 0.5332 to 0.6945; a paired re-evaluation of saved checkpoints, using 11 corrupted conditions and five realisations each, instead obtained 0.5370 versus 0.6374 (difference 0.1004). The subject-cluster bootstrap interval, conditional on the same four subjects, was [0.0683, 0.1283], but the exact two-sided subject-level prediction-swap test gave p=0.125 (four independent test subjects), so statistical significance is not established. Five independently generated subject-partition seeds (25 fully retrained folds) produced mean corrupted macro-F1 of 0.5375 for the fold-local RTPD reference versus 0.5449 for the validation-gated policy (mean difference 0.0074); fine-tuning was accepted in 5 of 25 folds. These results demonstrate that the validation gate usually retains the reference and that the observed adaptation gains are not established as broadly useful; they also indicate that robustness benefits are heterogeneous and that larger independent-subject evaluations are necessary.

Sensors doi: 10.3390/s26216969

Authors: Ivan Ortiz-Garces Fernando Molina-Granja Rommel Gutierrez Fernando Vela-Espin

Intrusion detectors for Internet of Things (IoT) systems are usually trained per device or site, so a verification layer must combine alerts from detectors built on different populations and base rates. This study compares two ways of turning such alerts into a decision: majority voting with a fixed threshold, and score aggregation with a threshold learned on labeled calibration data. Seven detectors trained separately on TON_IoT device populations voted on 56,000 shared events, and each rule was compared with the detector local to the event through an amplification factor A. Fixed-threshold majority voting amplified error in the point estimate (A=1.12 for 1−MCC), whereas learned-threshold aggregation reduced it (A≈0.62); device-level intervals, however, included neutrality. An equal-weight score with a learned threshold matched reliability-, margin-, and hybrid-weighted rules, so the improvement came from threshold placement rather than from weighting. In this setting, about 500 labeled target events sufficed to place the threshold, whereas source-domain reliability estimates were biased by up to 0.22 although their 95% sampling half-width was at most 0.011. Gains concentrated in devices with weak local detectors and shrank under natural prevalence. Under degraded validators, fixed-threshold voting was the most fragile rule. A synthetic vote layer overstated amplification, and a federated detector showed a shared threshold collapsing under domain shift.

Sensors doi: 10.3390/s26216968

Authors: Paul Reber Philipp Kersten Sören Majcherek Sören Hirsch

Moisture limits timber durability, and wetting begins at the surface. An interdigital capacitor (IDC) was printed by direct ink writing (DIW) of a silver paste onto beech wood, the substrate and sensing dielectric. The electrodes were designed within the smallest printable line width (210 μm) with an electrostatic finite element method (FEM) model; 90% of the sensitivity lies within 0.2 mm of the surface. One 24-finger sensor (width and spacing 300 μm) was characterized at 1–100 kHz over a 15-day sorption cycle (30–90% relative humidity (RH), 23 °C). Reference capacitors revealed a meter gain error of 22% at 100 kHz; after calibration, an identically printed glass reference agreed with the two-dimensional model within 3% (same session). The calibrated beech capacitance at 100 kHz rose from 7.0 to 13.7 pF, by 0.11 pF per percentage point of RH (short-term, noise-limited 3σ resolution 0.6–3.5% RH). Referred to moisture content (MC) assigned from tabulated equilibrium values, this is 0.47 pF per percentage point, close to the predicted 0.48 pF; the absolute values lie 5–17% below the model. The step response is two-phase. No gravimetric calibration was performed; the RH response is the primary result of this single-sensor study.

Sensors doi: 10.3390/s26216967

Authors: June Gyu Park Kyuman Cho

Four photodetectors, two in the balanced homodyne detector for in-phase (I) channel measurements and two in that for quadrature (Q) channel measurements, are needed for a conventional homodyne I/Q interferometer. Here, we present a new homodyne I/Q interferometer scheme for which two photodetectors, one for I-channel and the other for Q-channel measurements, are used for I/Q demodulation of the interference signal. The DC bias terms on the output signal from each photodetector are rejected by taking the difference between the signal and the temporally shifted signal. Then the induced phase on the interference signal can be measured by using the algorithm for I/Q phase demodulation. It will be shown that the sensitivity is limited by the corresponding photon noise. We will also show that in the vicinity of the turning points of a periodic modulation, where the phase difference may become smaller than the corresponding photon noise, or in the region where there is no measurable phase change the phase can be calculated through temporal averaging, which not only removes high-frequency noises but also increases the time delay for obtaining the difference signal.

Sensors doi: 10.3390/s26216966

Authors: Naqash Nasir Laura Cercenelli Stefano Stradiotti Michele Fiore Marta Bortoli Fabrizio Ferretti Roy Eagleson Brent A. Lanting Achille Tarsitano Massimiliano De Paolis Emanuela Marcelli

Augmented reality (AR) represents an emerging technology that is increasingly being integrated into surgical navigation across various medical disciplines, including orthopedic surgery. This preclinical study investigated the application of AR in orthopedic oncology and performed a comparative assessment of two head-mounted displays (HMDs): Microsoft HoloLens 2 and Magic Leap 2. The study evaluated the feasibility of AR-assisted tracing of planned osteotomies for bone tumor resections in orthopedic surgery. The objective was to evaluate the visualization capabilities, spatial guidance performance, and user interaction provided by the two AR platforms under controlled preclinical conditions. Accuracy of AR-guided tracing of planned osteotomies was quantified using custom cutting guides incorporating grooves of 1 mm, 2 mm, and 3 mm to measure deviations from the planned osteotomy lines. Both AR HMDs were tested under identical experimental conditions, and their performance and accuracy were comparatively analyzed. Magic Leap 2 demonstrated a trend toward greater accuracy in AR-guided tracing of osteotomies, achieving 82%, 95%, and 100% accuracy within 1 mm, 2 mm, and 3 mm thresholds of the planned osteotomy lines, respectively. In comparison, the HoloLens 2 achieved 65%, 85%, and 95% accuracy within the same respective thresholds. Ergonomic assessment also suggested favorable user comfort and handling characteristics for Magic Leap 2. Further studies involving larger cohorts and comparisons with established navigation techniques are required to confirm the potential advantages of HMD-based AR guidance for bone tumor resections and other orthopedic procedures.

Sensors doi: 10.3390/s26216965

Authors: Trent Piccicacco Brendyn Appleby Dale W. Chapman Simon Rosalie Kevin J. Netto Ritu Gupta

Analyzing patterns of play in team sports remains challenging despite advances in GPS tracking technology. While individual player statistics are well-documented, understanding what makes specific team plays successful requires extracting meaningful features from spatial–temporal data. This study addresses the question: can machine learning algorithms effectively characterize successful versus unsuccessful attacking plays in field hockey based on GPS-derived team movement patterns? We analyzed GPS data from a national men’s field hockey team recorded during the 2018 Men’s Hockey World Cup organized by the International Hockey Federation (FIH). We developed novel team-based features from longitude and latitude coordinates, including player closeness (~8–15 m Inter-Player spacing), corner positioning, movement energy via discrete wavelet transforms, and velocity–acceleration metrics. Thirteen features describing 63 attacking plays (42 successful, 21 unsuccessful) were classified using Random Forest (RF). Success was defined as achieving a goal, shot on goal, or attacking circle entry. Using 5-fold cross-validation repeated 10 times, the Random Forest model achieved a mean classification accuracy of 78% (95% CI: 75–81%) and an AUC of 0.88 (95%CI: 0.85–0.91). Sensitivity analyses showed that removing Opposition had minimal impact on predictive performance (accuracy 77.1%, AUC 0.862), while leave-one-match-out cross-validation yielded variable performance across held-out matches (accuracy 60–82%, AUC 0.69–0.96). The most discriminative features were player closeness during plays, high maximum acceleration (>2.5 m·s−2), number of athletes involved (9+ players), and mean distance traveled (~10 m). Play duration was also critical, with successful plays lasting 3–6 s. Machine learning can effectively characterize successful field hockey plays from GPS data when appropriate team-oriented features are extracted. Although the novel wavelet-based player energy metric showed a negative predictive contribution in its current form, it may still show promise for future load management applications. This computational methodological framework is potentially generalizable to other team sports requiring spatial–temporal pattern recognition.

Sensors doi: 10.3390/s26216962

Authors: Can Zhang Xingzhou Yan Lingen Luo Gehao Sheng

The UHF received signal strength indicator (RSSI) fingerprint map method was recently proposed to overcome the restrictions of the time delay algorithm widely used for partial discharge (PD) localization. However, existing studies focus on two-dimensional PD localization, while many situations require three-dimensional information. In addition, PD localization accuracy is highly dependent on the precision of prebuilt fingerprint maps, which remains challenging to achieve. Therefore, this paper proposes a hybrid three-dimensional PD localization method that combines RSSI fingerprint maps and a range-based algorithm. First, limited test points on several planes were selected to build RSSI fingerprint maps. Then, single-plane smooth interpolation was employed for each plane to obtain smooth three-dimensional fingerprint maps. Here, the limited test points compensate for the non-line-of-sight (NLOS) effect, while the interpolation method reduced the workload required to build the fingerprint map. Thirdly, enhanced four-dimensional fingerprint maps were established based on the Kriging method. Finally, a minimal deviation model was proposed to locate the position of the PD source by analyzing the relative ratio of the RSSI values across different sensors. Laboratory validation shows that the average error of the proposed method is 0.69 m, which demonstrates its accuracy and practical application value of the proposed method.

Sensors doi: 10.3390/s26216964

Authors: Zemiao Fang Yungang Hao Shuoyu Wang Tao Liu

During rehabilitation training, real-time monitoring and accurate identification of whether spasticity is occurring at the moment are crucial. This serves as the prerequisite for rehabilitation devices or robots to promptly implement interventions to alleviate spasticity, ensure patient safety, and ensure therapeutic efficacy. However, research on the real-time identification of the moment when the spasticity occurs by rehabilitation robots during training remains limited. In this paper, through clinical manifestations of spasticity and theoretical analysis, an index based on the divergence of acceleration and external torque changes (IDATC) and an IDATC-based method for identifying spasticity onset are proposed. A combined experiment was conducted to collect the kinematic, interaction force, and electromyography (EMG) data of stroke patients during spasticity provocation. The proposed spasticity identification method was validated using upper limb spasticity data collected from stroke patients. Preliminary results of the experiment demonstrate that the negative peak of the IDATC is highly consistent with the EMG peak of the corresponding spastic muscle, so it has potential to identify spasticity onset using the proposed IDATC-based spasticity identification method. Furthermore, provided that the threshold is appropriately selected, the proposed IDATC-based method may achieve an exceptionally high identification F1 score.

Sensors doi: 10.3390/s26216963

Authors: Zhiguo Zhang Wei Sun

Existing knowledge distillation methods for multimodal large language models (MLLMs) mainly align output distributions or intermediate representations, but rarely supervise how predictions depend on local visual evidence. This limitation may hinder the transfer of fine-grained perception and compositional reasoning. We propose Causal and Multi-source Visual Attribution Distillation (CMVAD), which transfers teacher-derived visual evidence during distillation. Region-level interventions first estimate token-level causal visual attributions, identifying image regions that influence each generated token. The Causal Evidence Adapter Fetch (CEAF) module jointly models token importance and semantic interactions to adaptively aggregate multi-source causal evidence into a unified representation. A Multi-granularity Attribution Distillation (MAD) objective transfers this representation via global evidence distribution alignment, critical-region alignment, and evidence-ranking consistency. Experiments on general visual question answering and compositional reasoning benchmarks show that CMVAD outperforms existing MLLM distillation methods, achieving larger gains in compositional reasoning while retaining competitive visual question answering performance. Results across teacher–student architectures demonstrate its generalizability. We further evaluate VisDrone UAV object detection through overall comparisons, providing supplementary evidence of CMVAD’s applicability to object recognition and localization in UAV imagery.

Sensors doi: 10.3390/s26216961

Authors: Junhui Xia Ning Zhang Mingdong Chen Haixia Mei

The electronic nose system is susceptible to cross-sensitivity, noise, and baseline drift, making simultaneous qualitative identification and quantitative prediction of volatile organic compounds (VOCs) challenging. To address these issues, this study proposes a CNN–Transformer multitask learning model based on response-patch sequences. A multichannel activity curve is first constructed to locate response onset and extract fixed-length samples. The long response sequence is then divided into ordered patches using a sliding window. A CNN captures local dynamic patterns associated with signal rise, peak, and recovery, while learnable positional embeddings and a Transformer model the sequential order and long-range dependencies among patches. Task-specific squeeze-and-excitation attention branches adaptively enhance classification- and regression-related features, and an uncertainty-weighted loss balances the two learning objectives. With a patch length of 75 and a stride of 37, the model achieved an Accuracy of 0.9717, an F1-score of 0.9696, an MAE of 5.0789 ppm, and an R2 of 0.9467. Comparative and ablation experiments confirmed its superior overall performance and the effectiveness of the proposed modules. The method provides a feasible solution for the integrated identification and quantitative analysis of VOCs in electronic-nose applications.

Sensors doi: 10.3390/s26216960

Authors: Matteo Cioni Maria Stella Valle Laszlo Kovacs Daniela Vorazzo Caterina Sergi Gianfranco Bosco

Surface electromyographic (SEMG) patterns during robot-assisted walking are still poorly documented. This study aimed to characterize the timing and synergies of muscle activations of the lower extremities during locomotion assisted with an end-effector robot. Healthy subjects walked at 1.8 km/h under two modes: passive and active with a full body weight load. The robotic gait cycle (RGC) was identified by detecting the flexion-extension phases of the right lower limb. SEMG signals were recorded from eight lower limb muscles, time-normalized with respect to the RGCs and averaged across 10 RGCs. Non-negative Matrix Factorization (NMF) identified three basic waveforms (factors) explaining 90.6% of the variance in the dataset of SEMG waveforms corresponding to three distinct muscle synergies: rectus femoris–vastus lateralis–vastus medialis (extension phase), gastrocnemius medialis–soleus–peroneus longus (mid-flexion phase), and tibialis anterior–biceps femoris longus (transition between extension and early flexion phases). Cluster analysis on the cosine distance matrix of the NMF factor coefficients further validated the three distinct muscle groupings. Temporal activation patterns were seemingly preserved between passive and active conditions, suggesting similar underlying neuromuscular organization. This study supports the use of clinical gait analysis, particularly with SEMG, to assess neuromuscular patterns of neurological patients before, during, and after robotic rehabilitation.

Sensors doi: 10.3390/s26216959

Authors: Jeong-Hyeon Moon Jangbeom Lee Bongjun Choi Heeduk Park

Infrared image radiance combines surface and atmospheric thermal emission with solar contributions, making agreement in total radiance insufficient to assess individual components. This study presents a model-based decomposition method using two mid-wave infrared (MWIR) bands with contrasting atmospheric absorption and solar sensitivity. Viewing geometry recovered from imagery and terrain data is combined with meteorological inputs. Ground-target temperatures retrieved from the longer-wavelength band are held fixed in shorter-wavelength predictions, without further target fitting. Observations on two days test these predictions across different atmospheric and surface contributions. For forests at 280 and 1080 m, spanning modeled surface-dominated and path-dominated conditions, shorter-band radiance predicted using longer-band temperatures differs from observations by at most 0.056 W m−2 sr−1 (maximum absolute residual across 12 comparisons at six times). Independently retrieved long-wave infrared forest temperatures agree with MWIR estimates within 2.6 °C. In two comparison frames, using observation-constrained atmospheric inputs reduces the root-mean-square (RMS) difference between observed and predicted longer-band sky radiance to 1/10.9 and 1/15.4 of the values obtained with an unmodified standard atmosphere, with the same offset retained across profiles within each frame. Greenhouse observations extend the analysis to diffuse solar reflection and localized brightening consistent with specular reflection. These results demonstrate how cross-band comparisons can assess the modeled radiance composition of field imagery beyond agreement in total signal.

Sensors doi: 10.3390/s26216958

Authors: Juan Luis Soler-Fernández Angel Diéguez J. Daniel Prades Oscar Alonso

This paper presents an n-FET-only operational transconductance amplifier (OTA) fabricated in a flexible 0.6 µm IGZO TFT process. The architecture combines a resistive-load differential pair with a three-stage common-source amplification chain and avoids using weak-conduction TFT loads to set the high-side currents. Open-loop measurements across 10 dies show a median DC gain of 56.6 dB and an input-offset distribution of approximately −313 ± 419 mV. The unloaded open-loop output excursion reaches 3.18 V. In a unity-gain follower configuration, the median closed-loop −3 dB bandwidth is 3.7 kHz, and stable operation is observed without an explicit compensation capacitor. The cross-chip average supply current varies from approximately 10 µA to 16 µA over the input sweep, corresponding to approximately 33–53 µW at 3.3 V. These results demonstrate a low-power IGZO amplifier for low-frequency sensor interfaces, while input offset, load-driving capability and supply coupling remain important limitations of the present implementation.

Sensors doi: 10.3390/s26216956

Authors: Yu Xu Xizheng Li Yinghao Song Kun Zhang Yuyang Liu Yue Sun

Reliable in situ soil moisture measurements provide essential information for irrigation scheduling and agricultural soil–water management. Among the available measurement techniques, capacitive soil moisture sensors based on frequency domain reflectometry (FDR) are widely used because of their high measurement accuracy and suitability for in situ monitoring. However, the resonant frequency of an FDR capacitive soil moisture sensor is affected simultaneously by temperature, soil electrical conductivity (EC), and water content conditions, and independent fixed coefficient corrections cannot adequately describe this coupled relationship. To address this problem, we developed a time division multiplexed three-parameter FDR sensor that uses a composite electrode to sequentially acquire a water content-related resonant frequency, a conductivity-sensitive voltage, and a temperature-sensitive voltage at the same sensing location. A joint three-parameter calibration model was then established to map the three outputs to the reference volumetric water content determined by oven drying. Comparison of five candidate calibration models showed that the backpropagation neural network (BPNN) provided the best overall performance on the development data. A paired 10-seed analysis showed no stable predictive advantage for WOA initialization over standard BPNN, so WOA was retained only as a secondary robustness analysis. After the WOA-BPNN calibration model was fixed and deployed on the embedded platform, repeat measurements under 72 temperature–salinity–water content conditions within the same prepared sandy loam system yielded an RMSE of 2.086 percentage points. Within the tested calibration domain, the coordinated three-parameter hardware and joint calibration model reduced water content errors associated with temperature and salinity variation. These repeat measurements were conducted in the same prepared sandy loam system and, therefore, do not demonstrate generalization across other soil textures, field sites, independently prepared sample batches, or seasons.

Sensors doi: 10.3390/s26216957

Authors: Umar Hasan Muhammad Ali Martuza Turki G. Alghamdi

Reliable medical-image classification requires assessment of accuracy, calibration, and uncertainty under image degradation. We analyze the unit-prior Evidential Mean Squared Error (EMSE) objective with masked Kullback–Leibler (KL) regularization and evaluate four ResNet-18 classifiers for four-class blood-smear classification. We derive confidence–uncertainty bounds and a Brier-loss lower bound, and we prove that masked KL leaves a target-evidence direction along which the objective approaches an unattained zero infimum. Using five optimization seeds, we compare regularized evidential learning, its no-KL ablation, softmax, and 20-pass Monte Carlo dropout on 3242 external images across 12 corruption conditions. Mean clean accuracy exceeds 0.998 for all methods; softmax has the best mean clean Brier score, expected calibration error (ECE), area under the risk–coverage curve (AURC), and error-detection AUROC. Across corruptions, the no-KL ablation has the highest equal-weighted mean accuracy, while reliability rankings vary by corruption and metric. None of 390 paired comparisons is significant after Holm correction. These findings do not establish a general reliability benefit from masked KL and clarify the distinct roles of predictive confidence and evidential uncertainty.

Sensors doi: 10.3390/s26216955

Authors: Mingyu Kim Cholsa Kosal Hyeongjun Mun Heejip Moon Jorge Jimenez

Directional sensors are widely used to monitor moving targets, yet conventional placement methods often represent targets as spatial points or overlook the coupling among sensor location, orientation, and target travel direction. This paper develops a theoretically grounded and scalable framework for joint placement and orientation of directional sensors for stochastic directional-trajectory detection. Directed trajectories are represented by heading and signed perpendicular offset, preserving opposite travel directions along the same geometric path, and their stochastic occurrence is formulated through an intensity model that can be represented using a log-Gaussian Cox process. Sensor performance is evaluated at the trajectory level by incorporating directional field-of-view geometry, distance-dependent detection probability, sensing frequency, dwell time, and repeated sensing opportunities. Network performance is quantified through missed-trajectory intensity and a corresponding void-probability objective. To improve computational efficiency, we introduce a dwell-time-aware binary sensing surrogate with deployment-aware radius calibration and an omni-greedy algorithm that ranks locations using an omnidirectional surrogate before orientation refinement. Theoretical analysis provides bounds on sensing approximation error, greedy decision preservation, location–orientation approximation, and end-to-end deployment performance. Experiments using real maritime AIS and wildlife GPS trajectories show that omni-greedy reduces selection-stage candidate evaluations by roughly 18–21× while maintaining detection performance close to full directional greedy optimization; when candidate responses are cached, total planning time is dominated by shared precomputation and radius calibration.

Sensors doi: 10.3390/s26216954

Authors: Rajeev Kumar Biswajit Brahma Jitendra Singh Tamang Paolo Barsocchi Akash Kumar Bhoi

This work proposes a highly sensitive Surface Plasmon Resonance (SPR) biosensor for label-free refractive index (RI) and biological sensing applications based on a BK7 prism-coupled bimetal (Ag/Au) and Tantalum Diselenide (TaSe2) hybrid structure. The transfer matrix method (TMM) is used to examine the sensing performance of the suggested structure while accounting for the effects of various biological sensing media (SM) and plasmonic layer thickness. To attain optimal plasmonic coupling, the impact of Ag layer thickness on sensitivity and minimum reflectance (Rmin) is methodically examined. For the long RI range of 1.35–1.40, the maximum sensitivity is 206.32°/RIU at the Rmin value. Depending on the sensing layer, the Ag thickness ranges from 39 to 42 nm, offering better resonance properties for various cancer detection. Basal, HeLa, Jurkat, PC12, MM-231 and MCF-7 have remarkable Rmin values and the highest sensitivity of 260.49°/RIU, 280.13°/RIU, 290.24°/RIU 260.49°/RIU, 280.13°/RIU and 290.24°/RIU with an optimized Ag thickness of 42–39 nm among the biological layers under investigation, respectively. Moreover, the penetration depth (PD) of 192.94 nm and 181.5 nm is attained at RIs of SM 1.35 and 1.40. The suggested SPR biosensor is a promising platform for biosensing, disease diagnosis, and real-time monitoring applications because of its exceptional sensitivity, low resonance reflectance, and improved detection capability.

Sensors doi: 10.3390/s26216953

Authors: Fangxiang Tang Bing Tang Shenbo Liu Dongxue Zhao Lijun Tang Junlong Tang

4D mmWave radar-vision fusion is a low-cost and promising solution for Unmanned Surface Vehicles navigation. However, existing methods mainly focus on single tasks and suffer from insufficient radar feature extraction and inefficient multi-modal fusion. This paper proposes RVFNet, a 4D mmWave radar-visual fusion multi-task perception network. It designs the Radar Feature Enhancement Module (RFEM) and alternately stacks RFEM with conventional convolution blocks to construct the Radar Feature Extraction Network (RFENet), which solves the problem of insufficient extraction of sparse and irregular radar point cloud features and suppresses water-surface clutter; it constructs the Multi-modal Feature Adaptive Fusion Module (MFAFM) with content-guided attention. MFAFM generates dynamic modal weights and adaptively balances visual and radar information to alleviate missed detection and false detection in harsh scenarios. Compared with mainstream methods, experimental results on the WaterScenes dataset show that RVFNet requires only 3.3 GFLOPs while achieving an mAP50 of 79.0%, an mAP50--95 of 46.7%, an mIoU of 84.0% for object segmentation, and an mIoU of 98.8% for drivable-area segmentation, demonstrating superior overall perception performance. Cross-scenario experiments on the FloW-RI water-surface dataset and the nuScenes v1.0-trainval autonomous-driving dataset further verify the strong generalization capability of RVFNet, demonstrating its potential for deployment on embedded devices.

Sensors doi: 10.3390/s26216952

Authors: Bernadete Márcia Voichcoski Franco da Rosa Pedro Kowacs Geysa Peters Ricardo Sbalqueiro Rosângela Carmo Wesley Hummig Percy Nohama

Schizophrenia Spectrum Disorder, widely known as causing behavioral changes, deprives many professionals—who remain capable of working—their jobs. This study aimed to investigate the use of electroencephalography exam (EEG) combined with standardized mental effort sequence as a complementary tool for clinical diagnosis and to monitor the treatment response in individuals with schizophrenia and those who want to return to professional activities. EEG recordings were acquired before (IR_Initial Register) and after (FR_Final Register) the completion of a set of timed mental effort tasks in a group of 25 clinically stable individuals diagnosed with schizophrenia with positive symptoms and an 11-member healthy volunteer control group. Data from both groups were compared using temporal and spectral descriptors, and the results were analyzed statistically. The results revealed that, in the FR phase, the Vtotal parameter for the schizophrenia group was 7.89% lower, and the frequency parameter was 16.35% lower than in the healthy group. Notable differences were observed in the tolerance parameter for the duration of sustained cognitive effort: 57.9% of healthy volunteers sustained 100% of the effort duration, compared with 6.1% of volunteers with schizophrenia.

Sensors doi: 10.3390/s26216951

Authors: Mehmet Yavuz Yağcı

Industrial Control Systems (ICSs) are increasingly evolving from isolated infrastructures into connected platforms due to the integration of information technology, industrial communication networks, and data-driven monitoring services. Although this transformation improves operational visibility and efficiency, it also expands the attack surface of ICS environments. Conventional vulnerability assessment and penetration testing methods are often unsuitable for operational ICSs because of safety, availability, and cost constraints. Similarly, intrusion detection and mitigation mechanisms require realistic attack data for development and validation, but collecting such data from running industrial systems is risky and impractical. Realistic ICS testbeds provide controlled environments for analyzing cyber-attacks, observing their consequences, and validating security mechanisms without disrupting operational facilities. This study presents FALCON (Factory Automation Lab for Cybersecurity & Operational Networks), a cyber-physical ICS testbed that integrates a real physical production process, PLC-based control, HMI/SCADA supervision, monitoring, logging capabilities, and IT–OT network segmentation. The testbed is designed to represent key functions commonly found in industrial production systems, including sensing, actuation, robotic transfer, quality control, labeling, and sequential process execution. FALCON supports both network-oriented and process-oriented attack scenarios within a controlled environment. Its layered architecture enables attacks to be initiated from different system points and analyzed across communication, control, monitoring, and physical process layers. The proposed testbed was evaluated in terms of design requirements and attack scenario coverage. It also supports a broad set of attack scenarios and enables the observation of immediate or delayed physical consequences.

Sensors doi: 10.3390/s26216950

Authors: Hongming Zheng Yuanyuan Qu Guangling Zhang

The accurate assessment of lithium-ion battery State of Health (SOH) is fundamental to the safe and reliable operation of battery management systems. The Belief Rule Base (BRB) integrates quantitative monitoring data and qualitative expert knowledge, thereby supporting SOH assessment. However, existing BRB models produce point-value outputs. The uncertainty of the assessment results cannot be quantified, nor can the belief level of the assessment be characterized. Moreover, information loss is caused by traditional evidential reasoning (ER) when interval belief degrees are processed. Interval reliability is not adequately addressed in existing optimization methods, as interval quality is not incorporated into the objective function. To address these problems, an interval SOH assessment method based on the Interval Belief Rule Base with Interval Evidential Reasoning (IBRB-IER) is proposed. Rule belief degrees are extended from point values to interval forms, enabling uncertainty quantification. Interval information is preserved throughout the fusion process by adopting interval evidential reasoning algorithms. A multi-objective composite fitness function is designed. Global parameter optimization is achieved through the projected covariance matrix adaptation evolution strategy (P-CMA-ES). SOH estimation intervals are finally produced as outputs. The effectiveness of the proposed method in SOH interval assessment is validated through experiments on the NASA dataset.

Sensors doi: 10.3390/s26216949

Authors: Hasan Mirzakhaninafchi Glen Rains Wesley Porter Canicius Mwitta Denis Kiobia Shekhar Thapa Peter Ngimbwa Christopher Ranger

Ethanol emitted from stressed woody tissues is a host-associated cue for ambrosia beetles, motivating investigation of portable ethanol-responsive sensors for future tree monitoring. This laboratory study evaluated temporal alignment between two asynchronously logged volatile organic compound (VOC)-responsive sensors during a denatured-alcohol chamber challenge. We selected the BACtrack Skyn wearable fuel-cell alcohol sensor for its compact form and timestamped output as a candidate for future tree-headspace measurements, and evaluated one device with an MQ-3 metal-oxide timing proxy and environmental logging in a closed static-headspace chamber. The dataset comprised five alcohol-free control runs (B0–B4) and three technical 1 μL denatured-alcohol challenge runs (EtOH1–EtOH3). Signals were processed on a common 1 min grid using a 10–15 min analytical baseline and a 15–45 min response-characterization window. MQ-3 peaks occurred at 17.0–17.3 min from the run origin (T0), whereas Skyn device-output peaks occurred at 18.8–19.2 min; peak-time offsets were 1.5–2.2 min. First-difference analysis selected a 2 min alignment setting on the primary 1 min grid in all three challenge runs; a supplementary 20 s analysis gave run-specific optima of 80–160 s. Across the three challenge runs, run-specific Pearson correlations ranged from 0.686 to 0.751 before alignment and from 0.861 to 0.930 after alignment; each coefficient was calculated from the paired 1 min observations in the 15–45 min response window. Normalized root-mean-square error (RMSE) ranged from 0.208 to 0.272 before alignment and from 0.121 to 0.230 after alignment, whereas rank correlation did not improve consistently. Alcohol-free controls remained below challenge-run peak magnitudes, although MQ-3 traces included drift and positive excursions. These within-run comparisons describe alignment in the tested configuration, not intrinsic sensor latency, concentration calibration, or validated tree-monitoring performance.

Sensors doi: 10.3390/s26216948

Authors: Hoang N. Bui Jose C. Matos Son N. Dang Nhung T. C. Nguyen

Field displacement measurements can support FE model updating, although the resulting model remains conditional on the available observations and modeling assumptions. This study proposes a hybrid Differential Evolution–Levenberg–Marquardt (DE–LM) framework for updating an FE model using train-induced displacement histories measured by linear variable differential transformers (LVDTs). A prestressed-concrete box-girder railway bridge was investigated, with the numerical model explicitly representing the box girder, sleepers, rails, and equivalent rail-fastening connections. Preliminary parameter screening was conducted to select three influential parameters: the box-girder elastic modulus and the effective vertical and lateral fastening stiffnesses. DE was employed to explore the bounded parameter space, after which LM refined the best global solution using local response sensitivities. Physical parameter bounds and regularisation were incorporated to limit non-unique and unrealistic solutions. Complete train passages were separated into calibration and independent validation datasets. The hybrid procedure reduced the initial objective value by approximately 90.5% and produced physically admissible parameter changes. For the independent validation dataset, the updated model achieved a mean normalized root mean square error of 2.17%, compared with 3.07% for DE-only, 4.94% for LM-only, and 7.46% for the initial model. The mean peak-displacement error was reduced to 0.70%. For the investigated case, the selected hybrid solution achieved lower validation errors than the reported DE-only and nominal-start LM solutions. The calibrated model provides a physics-based foundation for subsequent railway-bridge digital-twin development.

Sensors doi: 10.3390/s26216947

Authors: Gabriela Rozanski Jessica Polizzi Aidan Rogers Megha Amin Wilber Parada Iraheta Karolina Roszkowski Jenna Tosto-Mancuso David Putrino

Accurate assessment of impaired functional mobility is crucial to track recovery and guide interventions after stroke. Conventionally, physical therapists administer semi-objective rating scales or manually record performance on various tests. Technological developments, particularly automated algorithms that use motion data from inertial sensors, offer a promising solution for more efficient and objective evaluations. This cross-sectional analysis examined the agreement between therapist-recorded measures and a smartphone application (OneStep) for the Chair Stand Test (CST), Timed Up and Go (TUG), and the Six-Minute Walk Test (6MWT) in 42 participants with chronic stroke (mean age 54 years, 50% male). Analysis was measured via a two-way random-effect intraclass correlation coefficient analysis ICC(2,1) (absolute agreement) and Bland–Altman limits of agreement. ICC(2,1) was 0.989 [0.980, 0.994] for the CST, 0.995 [0.985, 0.998] for the TUG, and 0.921 [0.776, 0.967] for the 6MWT. Mean differences (therapist–app score) were −0.08 repetitions (CST), −1.03 s (TUG). and −69 feet (6MWT). Within-session repeatability was not detectably different between the app and therapist for the CST (SEM ratio app/therapist 0.91 [0.76, 1.05]) or the TUG (1.36 [0.92, 2.13]); the TUG estimates are imprecise and do not establish equivalence. The 6MWT was not considered in within-session repeatability as multiple tests were not completed due to patient fatigue. App failure rates were relatively low for the CST (2/80), TUG (3/76), and 6MWT (5/36). Correlations with patient-reported outcomes, clinician-rated outcomes, and gait metrics were exploratory (89 of 115 significant after false discovery rate correction). The results suggest close agreement for the CST, a small systematic difference for the TUG, and larger differences for the 6MWT. Future research should address the longitudinal predictive utility and broader implementation throughout the health care system of mobile applications like OneStep.

Sensors doi: 10.3390/s26216946

Authors: Lun Dong Han Yang Yong Zhang Shuai Wang Weiguang Li Chunting Zuo

Reliable crop mapping under incomplete seasonal observation is difficult because temporal evidence is sparse and public crop maps can disagree in parcel boundaries and class assignments. We investigate sparse-temporal crop mapping in Jilin Province using monthly Sentinel 1 and Sentinel 2 features acquired from June to September 2019, while treating three public crop maps as complementary supervision sources rather than independent ground truth. Their labels are harmonized into four map-derived classes and transformed into a hard pseudo target, heuristic pixel weights, soft label distributions, and agreement or conflict region codes. A full-temporal teacher then guides a sparse-temporal student through confidence-weighted losses, logit and feature distillation, and prototype regularization, allowing the same four-month interface to support different observation subsets. Under the fixed July and August setting, the STLRNet student reaches 0.792 OA, 0.742 Macro F1, and 0.596 mIoU, whereas complete June-to-September observations reach 0.821 OA. Soft pseudo labels improve OA from 0.721 to 0.751 relative to unweighted hard labels, and the student requires fewer parameters, operations, inference time, and memory than the full-temporal teacher. All metrics quantify agreement with the constructed pseudo reference, so the findings establish methodological feasibility for this Jilin 2019 case study rather than independent field accuracy or cross-region transfer.

Sensors doi: 10.3390/s26216945

Authors: Peng Gu Dong Gao Chunjiang Bian Minghui Zhu

When an unmanned underwater vehicle (UUV) with an aluminum-alloy hull maneuvers, its motion through the geomagnetic field generates an eddy-current magnetic field that interferes with the onboard magnetometer. We propose a method that extracts a low-dimensional spatial basis offline and reconstructs the eddy-current field within a local region online with two magnetometers. Offline, transient finite element method (FEM) simulations with unit excitations along the three body axes provide a quasi-static mode and a relaxation mode per excitation channel together with their step responses. Online, the time derivative of the primary magnetometer output drives a feed-forward reconstruction through the step responses, and the differential measurement of a single auxiliary magnetometer gives one amplitude factor, estimated sample by sample, that corrects the amplitude deviation of the offline model. The method requires neither angular-rate measurements nor calibration maneuvers. Six spatial modes represent the field of yaw, pitch, and roll maneuvers over the entire sensor candidate region with an error below 0.8%, and that of a combined reversing maneuver exciting all three channels with an error of 1.2%. With five classes of magnetometer error and two classes of model error injected, the region reconstruction error is 2.7–3.2% under a favorable installation and 6.1–7.3% under a typical installation. The residual at the primary magnetometer, the eddy-current interference remaining after vector compensation, is 8–9 nT and 14–15 nT. Open-loop feed-forward reconstruction yields 9.7–10.9%, and the complete Tolles–Lawson model calibrated on six other maneuvers yields 9–10 nT and 13–14 nT. In a further test, an independent FEM model with a 10% lower conductivity, an added conductor, and a new mesh generates the data, while the method keeps the original offline model. The amplitude factor converges to 0.90, the true ratio between the fields of the two models, and the residual stays within 4–10 nT. After the primary magnetometer is relocated, the residual without recalibration is 10 nT for the proposed method and 25–40 nT for the Tolles–Lawson model calibrated at the original position.

Sensors doi: 10.3390/s26216943

Authors: Mustafa Eriş Fatma Güneş Eriş Orhan Yaman

Recognizing affective states from electroencephalography (EEG) is a core task in affective computing, with applications in mental health monitoring and public safety. Most existing approaches rely on handcrafted features or image-based transformations of EEG, while transformer models capture global context at high computational cost. This study proposes KimiFormer, an end-to-end hybrid-attention Conformer that classifies raw EEG directly. A hierarchical Spatio-Temporal Stemming (STS) module tokenizes the multichannel signal and learns inter-electrode relationships. Stacked Conformer blocks combine linear and full attention in a 3:1 ratio, reducing computational cost while preserving long-range modeling. The model was evaluated on the public EEG Violence Dataset, recorded with a 14-channel wearable headset from 14 participants viewing violent and neutral videos. Under record-wise leave-one-record-out cross-validation with nested selection of the training hyperparameters, KimiFormer achieved 95.5% accuracy (97.0% under 10-fold cross-validation). On the external DREAMER dataset, its subject-independent performance remained close to chance (balanced accuracy 52.7% for valence and 50.2% for arousal), as did that of two established raw-EEG models. Ablation showed the STS module to be essential (65% accuracy without it). With 3.2 M parameters, 1.51 GFLOPs and an inference latency of 15.64 ms per sample, KimiFormer is suited to real-time, on-device processing of low-channel wearable EEG.

Sensors doi: 10.3390/s26216944

Authors: Junho Gong Kyuwan Lee Dooyong Cho

Displacement is a key response index for evaluating the serviceability and safety of civil infrastructure. Although vision sensors have emerged as non-contact and cost-effective alternatives to conventional sensing systems, their performance deteriorates under low-illumination and long-distance conditions because reduced spatial resolution and weakened local intensity information make feature extraction and tracking unreliable. To address this limitation, this study investigates the feasibility of integrating a real-world super-resolution (SR) framework into a vision-sensor-based system for dynamic displacement measurement under low-light conditions. A GAN-based RealSR model was employed to restore low-resolution image sequences, and the restored images were processed using a KAZE-based feature tracking approach. The proposed framework was validated through shaking table tests of a three-story frame structure under a low-illumination condition of 120 lux at measurement distances ranging from 10 to 40 m. Dynamic displacement responses induced by a harmonic excitation of 9.0 Hz were estimated and compared with reference measurements obtained from a linear variable differential transformer (LVDT). The results show that the SR-restored images provided more stable feature extraction and more reliable displacement tracking than bicubic upsampling, particularly for natural targets under low-light and long-range conditions. Furthermore, the application of SR improved the measurability of targets at extended distances, demonstrating that real-world SR is an effective preprocessing strategy for robust vision-based displacement measurement under low illumination.

Sensors doi: 10.3390/s26216942

Authors: Leqiang Zhong Wenzhi Han Yuhui Feng Qiaochu Yang Zhi Liu Peng Zhang

Multispectral disaster recognition draws on common optical measurements, but differences in feature construction, target definitions and evaluation design complicate comparisons across tasks. We present a unified feature framework that applies the same six-band interface, 78-dimensional spectral–spatial descriptor and training-side learning procedure to independently fitted flood, burned area and landslide experts. Spectral summaries, local variability and gradients describe each window; paired date and source auxiliary inputs provide task-specific extensions. Evaluation covered 88 flood patches from 11 events, 656 burned area windows in 23 source/overlap groups, and a prespecified 100-patch subset of the official Landslide4Sense test partition. The shared representation achieved F1 values of 0.8504, 0.8485 and 0.8169, respectively, under training-side classifier selection. The burned area temporal extension increased F1 to 0.8997, with gains over matched support post-event descriptors under both inner-selected and fixed Random Forest evaluation. The landslide auxiliary extension produced a larger gain under classifier selection than under fixed Random Forest, identifying an interaction between representation and classifier. Training label controls and negative class metrics distinguished learned discrimination from positive class prevalence. The framework provides a reproducible basis for testing shared optical descriptors and task-specific information under the distinct evaluation requirements of each task.

Sensors doi: 10.3390/s26216941

Authors: Ming Zhu

Ultrasonic sensing supports nondestructive testing, material characterization, motion and flow measurement, and biomedical imaging [...]

Sensors doi: 10.3390/s26216940

Authors: Sek Joo Shan Mazen Farid

With the rapid growth of motor vehicle ownership across cities, parking lots attached to large commercial supermarkets and transportation hubs generally face the common pain points of delayed vacant parking spot scheduling and low manual inspection efficiency. The market urgently needs an intelligent parking management system that can quickly cover an entire site and report real-time updates of parking space status. To address this gap, this paper presents an integrated system for visual parking occupancy detection based on drone aerial imagery, combining an oriented-bounding-box vehicle detector, automatic parking-slot discovery and tracking, and a dual-role web platform for managers and customers. Rather than proposing a new detection algorithm, the contribution of this work lies in the automatic, annotation-free slot inference, tracking and gap-filling mechanism and in the end-to-end system integration that removes the need for manual slot mapping or per-site calibration. The system was field tested on three real-world parking lots of increasing size and density and achieved an occupancy-classification accuracy of 82.4% and F1 of 0.805 (detection F1 = 0.827 and mAP@0.5 = 0.839) against manually annotated ground truth on 51 annotated frames, with a slot-discovery coverage (Slot Recall) of 0.833; these independently evaluated figures are prioritised over, as well as considerably lower than, the system’s own temporal-consistency detection scores of 98.7% accuracy/98.8% F1, which measures prediction stability without manual annotation. An itemised cost breakdown for a representative 10,000 m2 (~112-bay) lot indicates a deployment cost of RM6409 for the camera/UAV-based approach versus RM29,973 for an equivalent geomagnetic-sensor solution, which is approximately 21.4% of the geomagnetic up-front total (purchase and installation only); recurring UAV operating costs for continuous monitoring are discussed separately. The proposed approach offers a low-cost, sensor-free path toward intelligent parking management for large-scale lots.

Sensors doi: 10.3390/s26216934

Authors: Jiwon Jeong Seungjun Ahn

Restricted viewpoints on construction sites often yield partial point clouds in which weld seams are partly or entirely occluded. This study proposes a non-learning BIM-guided method that localizes BIM-defined nominal weld seams by registering a known H-section beam–column joint model to the visible flange and web surfaces. Given an isolated point cloud and the acquisition side, finite analytic member surfaces and a BIM-derived Hesse atlas support multiple admissible initial poses. Each pose is refined using the same multiscale robust point-to-surface objective, and a deterministic terminal criterion selects the pose used to transfer nominal seam paths. Neither seam labels nor reconstruction of occluded geometry is required. On 402 synthetic observations, median seam RMSE, ADD-RMS, and relative rotation error were 1.110 mm, 1.249 mm, and 0.067 degrees, respectively; 89.05% had seam RMSE within 10 mm. On 180 laboratory observations, median point-to-surface distance and 5 mm attachment were 2.137 mm and 82.04%. Median warm processing times on a matched panel were 2.52 s and 3.18 s for synthetic and laboratory inputs, respectively. These findings support the applicability of the proposed method to nominal weld-seam localization under restricted partial-view conditions relevant to construction-site welding.

Sensors doi: 10.3390/s26216939

Authors: Seongmun Jeong Sangwon Park Wontae Choe Chang Hyun Park Dahun Jung Jeesu Kim Hyung-Hoi Kim Chang-Seok Kim Pilwon Hur Hwidon Lee

Continuous knee flexion monitoring is important for rehabilitation and gait analysis, but optical motion capture is generally confined to laboratories, and wearable electrical sensors may be affected by drift, misalignment, or electromagnetic interference. This study presents a brace-integrated system based on a fiber Bragg grating (FBG) embedded in a functionally segmented, multi-hardness silicone strain transfer substrate. The substrate was placed in a dedicated brace recess, avoiding a separate rigid transduction mechanism. A portable interrogation unit enabled wireless transmission and real-time brace-derived angle estimation during sit-to-stand and walking tasks. Calibration against brace hinge angles from 0° to 90° showed a nonlinear Bragg wavelength–angle relationship, modeled using a fourth-order polynomial whose order was selected using a separately acquired validation set that was also used for performance assessment. Validation based on 10 angle-wise mean estimates yielded R2 = 0.9992, RMSE = 0.90°, and MAE = 0.68°, with Bland–Altman 95% limits of agreement of −1.46° to +2.05°. The 0° stability record had a wavelength standard deviation of 1.41 pm and a linear trend of −0.193 pm/min, and Allan deviation analysis gave a wavelength noise floor of 0.089 pm at an averaging time of 24 s. Noise-equivalent angular variation ranged from 0.08° to 0.52°. During a sustained 30° hold, the estimated angle drifted at −0.089°/min, but comparison of the 0° and 30° records in wavelength units showed that the undeformed condition reproduced approximately 59% of that drift, although the non-simultaneous comparison does not establish a common-mode thermal contribution. An uncompensated temperature change of 1 K corresponds to between 0.31° and 1.98° of apparent angle change. Proof-of-concept measurements in two participants, both of whom are authors, produced brace-derived profiles qualitatively consistent with expected task kinematics but revealed a cumulative zero-offset shift of 3.2° across two sit-to-stand cycles. Because no simultaneous anatomical reference was acquired, the dynamic results establish feasibility and waveform plausibility rather than measurement accuracy. The results support the engineering feasibility of this brace-integrated optical architecture, while identifying temperature compensation and cyclic loading characterization as prerequisites for anatomical and clinical validation.