Search Results (10,644)

In SAE Level 3 automated driving, drivers are required to supervise vehicle operation and respond to overtaking requests. Owing to task monotony and insufficient workload, drivers are prone to passive fatigue, which may impair vigilance and safety. This study investigated passive fatigue development during automated driving and proposed a multimodal detection method. Thirty licensed participants completed one automated driving task and one manual driving task in a driving simulator. Eye-movement and ECG/heart rate variability indicators were synchronously collected, and fatigue states were assessed using the Karolinska Sleepiness Scale. Results show that passive fatigue during automated driving developed differently from active fatigue during manual driving. Based on PERCLOS, pupil diameter, pupil diameter variation, SDNN, and LF/HF, an SVM-based passive fatigue detection model was developed to classify alert and passive fatigue states. Across repeated subject-wise validation, the model achieved an accuracy of 89.19%, sensitivity of 91.83%, specificity of 86.54%, balanced accuracy of 89.19%, F1 score of 89.48%, and precision of 87.28%, outperforming the model trained on manual driving active fatigue data. These findings demonstrate the need for scenario-specific driver-state monitoring models in automated driving systems and provide an applied physiological sensing approach for passive fatigue detection and warning design. Full article

Survival-rule induction is commonly guided by log-rank separation, whereas some prognostic questions target conditional future lifetime. We propose a directional finite-horizon mean residual life (MRL) criterion for survival-rule induction under right censoring, combining normalized subgroup support with a survival-weighted restricted-MRL discrepancy. The framework includes favorable and adverse objectives, training-only horizon selection and tuning, overlapping-rule prediction, and finite-candidate plug-in consistency. Across nine simulation scenarios (200 replications each), MRL recovered the true subgroup partition more accurately under delayed benefit, crossing hazards, delayed benefit with 60% censoring, and small-sample crossing; log-rank was stronger under proportional hazards, early-only effects, rare subgroups, and the adverse stress test. In repeated nested analyses, Cox proportional hazards achieved the lowest mean integrated Brier score (IBS) in WHAS100 (0.1797) and malignant melanoma (0.1300); controlled log-rank also yielded lower IBS than MRL (0.1926 vs. 0.2038 and 0.1358 vs. 0.1404, respectively). MRL nevertheless identified different conditional-lifetime structures. These results position MRL-guided rules as an estimand-specific complement to hazard-oriented methods rather than a universally superior predictor. Full article

To assess the long-term fatigue safety of fully enclosed high-speed railway noise barriers and to optimize maintenance strategies, a time-dependent probabilistic reliability framework focusing on fatigue damage and bolt preload relaxation is developed. The stochastic nature of train velocity, load scaling factor, and daily traffic volume is explicitly considered. Using the 350 km/h two-train passing scenario as the reference case, the stress time history at the column base is extracted. The damage per train pass is computed via rain-flow counting and Miner’s rule, and the time-dependent reliability indices are obtained through parallel Monte Carlo simulations (10⁴ samples, daily time steps over 50 years). The results show that the fatigue failure probability at 50 years is 0.11%, with a reliability index β ≈ 3.06 and a mean cumulative damage of 0.325; failures are concentrated in the 40–50-year period, and train velocity is identified as the most influential factor. Furthermore, 98.3% of the bolts require retightening within 50 years, with a median intervention time of 18.2 years—far earlier than the occurrence of fatigue failure. The joint system analysis reveals that the overall system failure is dominated by bolt relaxation; when the coupling effect is included, the fatigue failure probability increases to 0.18%. A phased maintenance strategy is accordingly proposed: monitor preload during years 0–15, perform comprehensive re-torquing during years 15–30, and intensify fatigue inspections during years 30–50. The proposed methodology provides a quantitative basis for life-cycle safety assessment and operational decision-making for fully enclosed high-speed railway noise barriers. Full article

Emergency Medical Services (EMS) personnel require exceptional spatial awareness and rapid decision-making within the confined environment of an ambulance. While Virtual Reality (VR) offers a safe alternative to traditional training, the lack of high-fidelity, regionally accurate, and VR-optimized 3D ambulance models limits its application. This study presents a scenario-driven methodology for developing a VR-ready 3D ambulance environment prototype tailored for Emergency Medical Technician (EMT) training. Utilizing reality-capture techniques, terrestrial laser scanning was performed to accurately document the interior of a standard Hungarian ambulance simulator. The resulting point cloud underwent systematic processing, manual retopology, PBR shading, and the implementation of a custom dual-rigging animation system to optimize complex mechanical movements—such as stretcher operations—for standalone VR platforms. The workflow successfully reduced the vertex count to 25,373 while maintaining millimeter-level spatial fidelity. Technical evaluation confirmed that geometrical, functional, and material objectives were fulfilled, whereas pedagogical implementation remains incomplete. Structural accuracy and animation readiness were verified through preliminary inspection within Blender’s VR viewport inspector. However, interactive game-engine integration remains future work, and educational effectiveness has not yet been tested with EMT learners. Overall, this workflow delivers a 3D asset foundation that establishes the necessary technical basis for subsequent software implementation and clinical evaluation. Full article

The large-scale integration of renewable energy causes grid strength to vary over a wide range, exposing grid-following (GFL) inverters to both mid- and high-frequency resonance and subsynchronous oscillation (SSO). Conventional fixed-parameter designs cannot simultaneously maintain stability and dynamic performance because the phase-locked loop (PLL) and grid-voltage feedforward (GVF) dominate different frequency bands. This paper therefore proposes a backpropagation-neural-network (BPNN)-based adaptive parameter control strategy with frequency-band-coordinated regulation. First, a q-axis small-signal output-admittance model incorporating the current loop, digital delay, PLL, and GVF is established. The model reveals that the GVF coefficient primarily shapes mid- and high-frequency admittance under strong and moderately weak grids, whereas the PLL bandwidth becomes the dominant factor in low-frequency and subsynchronous stability under ultra-weak grids. Based on this mechanism, a BPNN is constructed with the grid short-circuit ratio (SCR) as the input and the GVF coefficient and PLL bandwidth as the outputs. Training targets are generated offline using parameter sweeps and performance screening based on current total harmonic distortion, Point of Common Coupling (PCC) voltage error, and settling time. During operation, the GVF coefficient is adjusted first, and the PLL bandwidth is reduced only when the grid becomes ultra-weak. Simulation results over $\mathrm{SCR}=1.25$–10 demonstrate that the proposed strategy preserves stable operation while providing better transient and harmonic performance than fixed-parameter and single-parameter tuning schemes in the cases studied. Full article

Timely vehicular sensing is important for traffic monitoring, cooperative driving, and road-safety management. High mobility, time-varying wireless conditions, and limited edge resources nevertheless make information freshness difficult to maintain. This paper studies age of information (AoI) minimization in a three-layer UAV-assisted edge network comprising vehicle devices (VDs), unmanned aerial vehicles (UAVs), and a cloud center (CC). VDs periodically generate sensor-data packets, UAVs provide mobile edge processing and data-relaying services, and the CC coordinates system-wide resource allocation. The joint optimization of sensor-data transmission, UAV movement, packet processing, computation offloading, and bandwidth allocation is formulated within a cooperative multi-agent framework. To solve this problem, we propose a collaborative heterogeneous federated actor–critic (CHFAC) framework. Its significance-aware federated learning mechanism evaluates local model updates according to update significance, alignment with the global learning direction, and training stability and uses the resulting contribution scores for non-uniform agent selection and contribution-weighted aggregation. In the considered simulation setting, evaluation over 1000 test episodes yields an average AoI of $7.45\pm 1.65$ and a worst-case AoI of $38.72\pm 24.06$. The average AoI is $79.0\%$, $63.9\%$, and $29.2\%$ lower than that obtained by the implemented HF-MARL, H-MAAC, and non-federated baselines, respectively. These results demonstrate the effectiveness of CHFAC for freshness-aware vehicular sensing in dynamic UAV-assisted edge environments. Full article

We present a reward-free control framework for balancing and steering a two-wheeled scooter using a diffusion-based world model. Rather than engineering a reward, we specify goals directly in observation space: target values (e.g., zero roll and zero yaw error) are pinned through a continuous mask, and classifier-free guidance amplifies the goal signal during trajectory generation. Because the mask is continuous at inference, goals can be traded off online (for instance, relaxing the balance constraint during sharp turns to allow necessary leaning) without retraining. The model is a FiLM-Mixer denoising network trained with V-prediction diffusion. At deployment, the controller runs in real time using a single diffusion step with warm-started predictions. We validate the approach on a full-sized Thormang3 humanoid operating a Gogoro Viva scooter in simulation, and deploy it on physical hardware. It matches a PPO baseline tuned with six reward components on balance, survival, and heading tracking while producing smoother commands, all without the per-task reward-shaping step. Diffusion training introduces its own loss-weight hyperparameters; unlike reward weights, however, these are task-agnostic. They govern the denoising procedure rather than the desired behavior, and are therefore set once and reused unchanged across goals rather than re-tuned for each new task. Because the model learns to predict trajectories rather than to maximize a reward, its training signal depends only on observed states and actions, not on reward labels. Real hardware recordings can therefore be folded directly into the same loss, providing a route toward closing the sim-to-real gap that reward-based methods such as PPO structurally cannot use. Full article

Artificially replicating the extraordinary adaptive potential of organisms remains difficult. Machine learning approaches based on big data pursue behavioral adaptation through generalization from training data, but often learn slowly and struggle with out-of-distribution situations. We propose that organisms may not adapt primarily through generalization alone but by rapidly eliminating infeasible solutions through online trial-and-error, effectively performing a search over the behavior space. If this search is complete, it is guaranteed to find an existing physically robust solution within a finite but unbounded time. For continuous behavioral domains that contain uncountably infinite behaviors, we introduce a mathematical framework for constructing a countably infinite dense subset of all behaviors using a mutable graph to segment behavior space, allowing any behavior to be progressively approximated arbitrarily well. Graph evolution is regulated by a heuristic feedback loop between outward growth and internal refinement; refinement is partially determined by a Bernoulli variance term related to binary entropy. Using this construction, we implement a proof-of-concept behavioral search algorithm and evaluate it on maze navigation and simple continuous control tasks. These preliminary results establish the practical feasibility of this approach in low-dimensional simulated environments while exposing unresolved limitations in terms of dimensional scaling and the incorporation of prior information. Full article

Multiple-criteria decision analysis (MCDA) is a branch of operations research concerned with choices based on several quantifiable factors. Here, we focus on high-dimensional markets, in which each available product is described by a large number of features. Taking a probabilistic approach, we define a market as a collection of alternatives in a decision-making scenario governed by a linear utility function. Analytic approximations for the market share and its moments are derived in the limit of a large population and a large number of measured features. We identify a single parameter, termed the degree of subjectivity, that places markets on a continuous spectrum ranging from fully objective to fully subjective. At an intermediate value, the market is competitive in the sense that it maximizes the entropy of the market-share distribution. Empirical analysis of several real markets indicates that they can indeed be classified by this parameter, yielding predictable decision patterns and a unified, relative measure of competitiveness across markets. Simulations involving non-linear utility functions and a trained machine-learning classifier provide preliminary evidence that similar behavior may also arise beyond the linear model, suggesting that the degree of subjectivity may be useful as a diagnostic in some broader multi-feature decision problems. Full article

Reliable attribution of artificial intelligence (AI)-generated text to a specific large language model (LLM) matters increasingly as LLMs proliferate, yet where quantum machine learning actually stands on this task has, to our knowledge, never been measured systematically. We benchmark the quantum support vector machine (QSVM) for binary attribution between Gemma 3 and Qwen 2.5 on a 5800-sample corpus from paired prompts: 83 configurations sweeping qubit count, regularization, training-set size, feature-map family, and circuit depth under exact, noiseless classical statevector simulation. QSVM validation accuracy plateaus at approximately 88%, whereas a classical support vector machine with a radial basis function kernel reaches approximately 97.8% on the identical fourteen-dimensional inputs: the ceiling belongs to the quantum (fidelity) kernel, not to the input representation. We measure the mechanism: off-diagonal quantum kernel values shrink exponentially with qubit count, the signature of exponential kernel concentration. The same classical model recovers the stylometric attribution fingerprint, showing it belongs to the shared feature pipeline rather than to the quantum kernel. All large-scale headline results generalize to an independent 1000-text test set produced after every design decision was frozen. The study provides a cautionary, reproducible benchmark for quantum kernel natural language processing and outlines an open-set extension as future work. Full article

Large-scale public events generate concentrated passenger demand during egress periods, often overwhelming urban transit systems. This paper proposes a multi-modal evacuation framework that coordinates in-service buses temporarily diverted from existing lines and dedicated shuttle vehicles pre-positioned at depots. The problem is formulated as a two-layer stochastic optimization under travel time uncertainty: the upper layer determines pre-event shuttle fleet sizing, while the lower layer makes real-time dispatching decisions for both modes. We propose an Uncertainty-Aware Reinforcement Learning framework with Categorical DQN (UARL-CD) that learns a robust dispatching policy through a reward function aligned with the lower-level objective, explicitly accounting for travel time uncertainty via distributional value representation and stochastic training, with an action masking mechanism enforcing operational constraints. Simulation experiments based on a realistic stadium evacuation scenario demonstrate that the proposed framework significantly outperforms deterministic optimization and rule-based strategies, achieving a 31.6% reduction in evacuation completion time and a 48.4% reduction in average passenger waiting time compared to shuttles alone, while maintaining robustness to travel time uncertainty with only 4.0% performance degradation and online decisions executed within the 2-min decision interval. Full article

This paper presents an experimental evaluation of a smart glasses-based cyber-physical system for the recognition of dangerous objects and the detection of dynamic life-threatening events. The proposed system integrates real-time computer vision and sound-based localization mechanisms in order to enhance situational awareness and user safety. The dangerous object recognition module is based on a deep learning model derived from the YOLO architecture and MobileNet classifiers, extended with distance estimation capabilities. The dynamic hazard detection module combines auditory localization using interaural time and level differences with visual confirmation through object detection. The experimental study was conducted using datasets of hazardous objects and simulated dynamic scenarios, including moving threats and sudden acoustic events. The dangerous object recognition module achieved F1-scores of up to 100.00% during training and 99.33% during validation, demonstrating high robustness and reliability. The dynamic hazard detection module achieved 0.95 training accuracy and 0.71 validation accuracy under dynamic environmental conditions. Comparative experiments on two hardware platforms showed that the ASUS ROG Zephyrus M16 reduced training time by up to seven times compared to the GIGABYTE GA-A320M-H platform without affecting recognition accuracy. The obtained results confirm the applicability of the proposed multimodal approach for real-time safety systems and wearable assistive technologies operating in complex environments. Full article

With the advancement of the new-type power system, the exponentially growing high-frequency distribution and consumption data imposes heavy transmission and processing pressure on resource-constrained edge devices. Existing compression methods face two core limitations: static algorithm configurations that fail to adapt to dynamic time-varying power load characteristics, and complex computations that are difficult to deploy on resource-constrained edge terminals. To address these issues, this paper proposes a cloud-edge collaborative adaptive compression method based on CNN-LSTM load forecasting and Q-learning decision-making. A three-layer “cloud-edge-terminal” architecture is built to decouple compression decision-making from edge execution. The cloud employs a hybrid one-dimensional CNN and single-layer LSTM (1D-CNN-LSTM) for high-precision short-term load forecasting, and establishes an adaptive Q-learning decision mechanism to issue differentiated compression instructions according to varying load characteristics. The edge terminals receive these instructions and perform lightweight lossless compression accordingly. Simulation results show that the CNN-LSTM model achieves a MAPE of 7.59%. The Q-learning agent converges to an average reward of 65.35% during training and achieves a 66.73% overall compression ratio on the unseen test set, outperforming the fixed LZW baseline by approximately 6 percentage points. Furthermore, the proposed method improves the edge processing throughput by approximately 6.5 to 10.4 times compared to the comparative baselines. These results suggest that the cloud-edge collaborative approach offers a promising direction for alleviating edge pressure and balancing compression efficiency with computational overhead in massive power data transmission scenarios. Full article

The stream of onions entering a self-propelled onion collector varies with field conditions, feeding density, and the transient lifting behavior of the onion–soil mass, whereas the collection conveyor is conventionally operated at a fixed speed, wasting hydraulic energy. This study proposes and evaluates, through field-calibrated simulation, a vision-based feedforward conveyor speed control framework that couples reel-synchronous onion counting with variable-displacement pump control. To mitigate the periodic occlusion caused by the compact dual-conveyor structure, a frame-selection method synchronized with the detected conveyor position was implemented, and a YOLOv8n detector was trained and evaluated on 314 images extracted from 32 indoor and field source videos partitioned at the source-video level. On the independent test subset, the model achieved precision, recall, and mAP@0.5 of 0.952, 0.944, and 0.959, and the proposed counting method maintained count recovery ratios above 95% across all engine speeds in 45 independent field trials, outperforming fixed-period sampling and tracking-based baselines by 6.6 and 3.8 percentage points, respectively. An AMESim model of the fixed-displacement hydraulic system was calibrated and shown to be consistent with field measurements and was then used to evaluate the variable-displacement configuration. Net conveyor-related fuel savings (engine no-load consumption subtracted) were estimated at 14.3–18.7% under the field-identified constant transmission efficiency, with conservative estimates of 11.5–16.9% when partial-displacement efficiency losses were accounted for through an anchored loss model. These results demonstrate the feasibility of vision-based feedforward conveyor speed control and its potential for energy savings in bulb-crop collection. Full article

This paper proposes a novel outpainting technique for quadri-wave lateral shearing interferograms which employs Quadri-Wave Generative Adversarial Networks (QW-GANs). QW-GAN is adversarially trained to capture the high-order statistics of real interferograms. It thereby avoids the Gibbs-ringing artifacts and spectral leakage caused by deterministic interpolation or single-frame extrapolation. The recovered fringes remain physically consistent with preserved Fourier support, enabling more accurate wavefront phase retrieval. Numerical simulations and comparisons with aberrations (defocus, astigmatism, coma, spherical, and random), along with real experimental results, demonstrate the enhanced reconstruction precision of our method, providing an efficient solution for wavefront reconstruction in optical applications. Full article