Search Results (2,148)

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 study explored whether participation in a digitally mediated sustainability literacy intervention was associated with short-term improvements in sustainability-related learning outcomes across two contrasting contexts: Portugal (stability) and Myanmar (polycrisis). Using a Most-Different Systems Design, 152 university students participated in a four-week Digital Sustainability Literacy Online Course, with food systems and responsible consumption providing a practical context for integrating environmental, social, economic, and governance dimensions of sustainability. A mixed-methods approach combined pre- and post-intervention assessments using the Sustainability Consciousness Questionnaire, operationalized through Knowledge, Attitudes, Self-reported Practices, and Self-reported Explanatory Efficacy (KAP + E), alongside engagement measures informed by ARCS and EGameFlow. Quantitative data were analyzed using paired- and independent-samples t-tests and mixed-design ANOVAs, while qualitative responses were examined using deductive thematic analysis. Significant pre–post improvements were observed across all four outcomes in both cohorts. Contextual differences were observed for knowledge and attitudes, while a significant Time × Context interaction for explanatory efficacy indicated greater gains in Portugal. Qualitatively, environmental and social dimensions predominated in both cohorts, while governance was more prominent in Portugal. Overall, participation was associated with short-term gains across multiple sustainability-related outcomes, while explanatory efficacy showed greater contextual sensitivity. Given the absence of a control group and complete confounding of the country with the stability/polycrisis condition, the findings should be interpreted as comparative associations rather than causal effects of the intervention or polycrisis. Full article

Early architectural decisions shape building energy demand and life cycle carbon, yet digital support remains fragmented across modelling, simulation and performance workflows. This systematic review examines how digital twin (DT) frameworks, tools and workflows are applied at the architectural design stage to support net-zero building performance. It investigates whether design-stage digital twins function as decision support systems or remain BIM-plus-simulation workflows carrying a twin label. Following PRISMA 2020, Scopus, Web of Science, IEEE Xplore, ScienceDirect, SpringerLink and Taylor & Francis Online were searched, supplemented by citation searching. Searches were completed on 28 March 2026 and restricted to English-language, peer-reviewed publications meeting eligibility criteria. Thirty-three studies formed the analysed corpus, comprising 10 Tier 1 core digital twin studies, 15 Tier 2 DT-oriented studies and 8 Tier 3 DT-enabling studies, while 19 review and contextual sources supported framing. Studies were coded by DT conceptualisation, framework type, enabling technology, life cycle stage, net-zero indicator and validation approach. Findings were synthesised descriptively and thematically, and evidence maturity was assessed using a six-domain appraisal covering reporting quality, digital twin completeness, design-stage relevance, validation quality, reproducibility and architect usability. Only 11 studies, representing 33% of the corpus, were anchored in concept or schematic design. BIM, building-performance simulation and parametric or generative modelling were dominant, while IoT and AI or machine learning supported prediction, surrogate modelling and control. Energy was addressed in 26 studies and thermal comfort in 10, whereas embodied carbon, daylight, renewable generation and indoor air quality received limited attention. Eight studies were classified in the ‘Measured/large empirical’ validation class. Methodological heterogeneity and limited empirical validation precluded meta-analysis. Design-stage digital twins remain emerging rather than mature decision support systems. The review was retrospectively registered on the Open Science Framework (DOI: 10.17605/OSF.IO/N5Z8H) and received no external funding. Full article

Urban rail transit systems face a pronounced contradiction between escalating traction energy consumption and passenger service quality under the carbon peaking and carbon neutrality goals. This study proposes a hybrid optimization framework offline combining deep reinforcement learning (DRL) and Non-dominated Sorting Genetic Algorithm III (NSGA-III) for multi-objective train timetable optimization in urban rail transit systems. The proposed model simultaneously considers energy consumption, passenger waiting time, and regenerative braking energy utilization. To address the limitations of traditional evolutionary and learning-based methods, a two-stage optimization architecture is developed. In the offline stage, NSGA-III is used to generate high-quality Pareto-optimal solutions, which are utilized to initialize the experience replay buffer of a Double Deep Q-Network (Double DQN). In the online stage, the DRL agent performs adaptive timetable adjustments under dynamic passenger demand. NSGA-III is used exclusively in the offline stage for Pareto solution generation and DQN pre-training; no NSGA-III optimization is performed during online execution. Experimental results on a real-world metro case study demonstrate that the proposed method achieves significant improvements in energy efficiency and service quality compared with baseline methods. The results confirm that the hybrid framework provides an effective and scalable solution for real-time metro timetable optimization problems. Full article

This study aims to explore how the functional characteristics of virtual models in autonomous driving systems influence users’ psychological needs, thereby shaping driving habits and continued usage intentions. As autonomous driving systems evolve toward learning-based intelligence, human–computer interaction interfaces play a crucial role in shaping driver behavior and sustained system adoption. However, prior research has primarily focused on trust and intention to use, with less attention paid to how motivational and system-feature factors jointly influence driving habits and continuous use intentions. Building upon self-determination theory (SDT), this study constructs an extended framework for human–AI co-driving behavior, examining the impact of three intrinsic psychological drivers (perceived autonomy importance, self-efficacy, and identification) on driving habits and continuous use intentions, while also considering the visual factors of virtual models and nine technical feature factors. Using online questionnaires, 614 valid samples were collected and empirically tested using PLS-SEM and IPMA. Results indicate that the perceived importance of autonomy significantly positively influences both driving habits and continuous use intentions; self-efficacy, identification, and visual factors did not show significant effects at either stage. Among the technical feature factors, data acquisition and feedback, intelligent driving modes, and risk perception capabilities significantly enhanced driving habits and continuous use intentions, achieving an optimal “high importance–high performance” match in the IPMA matrix. IPMA further revealed that perceived autonomy importance was associated with a “high importance–low performance” mismatch in the driving habit formation stage, representing a critical shortfall requiring urgent optimization; intelligent driving modes and data acquisition and feedback, conversely, demonstrated both high importance and high performance in driving continuous use intentions. These findings extend the application of self-determination theory to autonomous driving scenarios into a unified framework integrating motivational and system-feature factors, providing empirical evidence for phased optimization of human–AI co-driving system design and evaluation. Full article

The space–air–ground integrated network (SAGIN) enhanced by mobile edge computing (MEC) has emerged as a promising architecture for future 6G systems, providing wide-area coverage and distributed computing capabilities. By representing requests as service function chains (SFCs), network function virtualization (NFV) enables coordinated orchestration of underlying resources. However, SFC orchestration in SAGIN-MEC faces three significant challenges, including multi-layer resource heterogeneity, topology dynamics, and complex sequential dependencies within SFCs. To address these challenges, this paper proposes GT-PPO, a deep reinforcement learning (DRL)-based approach for online SFC orchestration designed to maximize network profit while minimizing end-to-end (E2E) delay. GT-PPO employs a graph attention network (GAT) to identify interactions among heterogeneous nodes and extract rich feature information from the dynamic physical network. Additionally, it leverages the Transformer self-attention mechanism to encode the SFC context based on resource demands and current deployment progress, thereby capturing global dependencies among virtual network functions (VNFs). Extensive simulation results demonstrate that, under high-load conditions, GT-PPO outperforms representative baselines, increasing the request acceptance ratio and network profit by 5.62% and 16.71%, respectively, while reducing the average E2E delay by 15.46%. Full article

Background/Objectives: Dental implants are a reliable treatment for tooth loss, but identifying the implant brand when patient records are unavailable remains a clinical challenge that complicates prosthetic repair and complication management. This study aimed to develop and evaluate a deep learning-based system for automated identification of dental implant brands from panoramic and periapical radiographs. Methods: In this retrospective study, anonymized radiographs containing implants of twelve known brands were obtained from the archives of Necmettin Erbakan University Faculty of Dentistry. A two-stage pipeline was employed: a YOLOv11 detector first localized and cropped the implant regions, after which an EfficientNetV2-M convolutional neural network, fine-tuned via transfer learning, classified the implant brand. Class imbalance was addressed through offline and online data augmentation. Results: On the held-out test set of 531 implant crops spanning twelve brands, the classifier achieved an overall accuracy of 96.23% (95% CI 94.5–97.7%), a macro-averaged F1-score of 0.953, and a macro-averaged ROC-AUC of 0.991; the complete pipeline evaluated end to end on detector-predicted crops reached 96.0% match-conditional implant-level accuracy, corresponding to a precision-aware end-to-end identification F1-score of 88.2% (precision 81.6%, recall 96.0%) when all predicted boxes, including false detections, were counted. Grad-CAM analysis, including misclassified and low-confidence cases, indicated that predictions were based on clinically meaningful implant morphology. Conclusions: These findings indicate that the proposed two-stage approach provides accurate and interpretable implant brand identification, supporting its potential as a clinical decision support tool. Full article

The convergence of decoloniality, artificial intelligence (AI), and university classroom practice has emerged as a pressing concern in African higher education. While decolonial movements such as Rhodes Must Fall and #FeesMustFall have reignited debates regarding epistemic justice within African universities, the rapid mainstreaming of generative AI has introduced new inquiries concerning whose knowledge is valued, which languages are acknowledged, and whose pedagogies are perpetuated through machine learning systems. This systematic review maps the trends in studies that engage with decoloniality, AI, and university classrooms, focusing specifically on African universities and African scholarship from 2010 to 2025. Adhering to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2020 guidelines, the review examined Scopus, Web of Science, African Journals Online, Sabinet, ERIC, and Google Scholar, supplemented by citation chasing. A total of twenty-five studies of varying types and peer-review status met the inclusion criteria. The findings indicate a marked acceleration in scholarly output following 2020, the emergence of three dominant thematic clusters (epistemic injustice and curriculum, algorithmic coloniality and data extractivism, and pedagogical adaptation of generative AI), and a significant concentration of research in South Africa, alongside emerging Pan-African scholarship on African-language AI. The review concludes that African scholarship is formulating a coherent decolonial AI research agenda; however, the scarcity of empirical evidence from classroom settings, the marginalisation of African languages within AI corpora, and uneven regional representation hinder the advancement of the field. Implications are discussed for institutional policy, curriculum design, and future research directions. Full article

Owing to unknown hydrodynamic characteristics, unmeasurable external disturbances and uncertain control coefficients, conventional control strategies that rely on precise mechanistic models often suffer from notable performance degradation. To address this problem, a novel data-driven Nussbaum-RBF Sliding Mode Control (NRSMC) strategy is proposed for USV heading control. In the proposed framework, a Radial Basis Function (RBF) neural network is employed as a data-driven approximator to learn unknown nonlinear system dynamics online, avoiding the dependence on accurate prior mathematical models. Based on the online-learned dynamic information, a sliding mode control (SMC) mechanism is developed to enhance robustness against approximation errors and external disturbances, and a boundary layer technique is introduced to alleviate the chattering phenomenon. Furthermore, a Nussbaum function is incorporated to address the unknown control coefficients problem, ensuring system stability without requiring prior knowledge of the control coefficients. The stability of the closed-loop system is rigorously analyzed using Lyapunov theory. Comparative simulation results demonstrate that the proposed NRSMC strategy achieves superior tracking accuracy and robustness compared with the Nussbaum-RBF Adaptive Backstepping Control (NRABC) method. Moreover, field experiments conducted further validate the effectiveness and adaptability of the proposed data-driven control approach. Full article

Generalization under deployment shifts remains a major challenge for visuomotor robot manipulation, especially when changes in camera placement and control behavior are unknown to the policy. This paper introduces the System-Context-Conditioned Diffusion Policy (SCE-DP), a history-conditioned manipulation policy designed to adapt its latent context to simulated changes in camera extrinsics, action scale, and bounded control delay without explicit recalibration or test-time model-weight updates. SCE-DP encodes a short history of issued commands and their subsequent visual and proprioceptive responses into a latent system context, which conditions the diffusion-based action denoising process. Auxiliary system-parameter regression and one-step response prediction further encourage the context to capture deployment-relevant system properties and their behavioral consequences. We evaluate SCE-DP on five independently trained ManiSkill manipulation tasks across held-out in-range configurations, a withheld camera-delay composition, and mild extrapolation settings. SCE-DP improves the five-task macro-average over domain-randomized Diffusion Policy by 14.5 percentage points on held-out configurations and 15.7 points on the withheld camera-delay composition. Across four shifted evaluation suites, it achieves an average success rate of 64.3%, compared with 48.9% for the domain-randomized baseline, while preserving nominal performance. Command-only, response-only, previous-action, recovery-source, and matched explicit-parameter controls indicate that the gain is not explained solely by command statistics, redundant command input, or recovery-data composition. These results show in simulation that command–response history is an effective source of online latent system context under coupled visual and control shifts; real-robot generalization remains to be established. Full article

Protected horticultural crops are commonly produced at high planting densities and have short production cycles; imbalances in water and fertilizer supply can rapidly affect plant vigor, yield, and quality. Non-destructive diagnostic methods are therefore needed to characterize plant nutritional status under greenhouse conditions. Spectral imaging can simultaneously capture spatial and spectral information associated with pigments, water status, tissue structure, and canopy phenotype. It does not directly detect nutrient ions; rather, it captures physiological and structural responses that may be associated with nutrient status and may also be influenced by water deficit, disease, temperature, salinity, phenology, and genotype. This review focuses on crops grown in soil, substrate, and hydroponic systems under greenhouse conditions. Studies conducted in vertical farms, growth chambers, and open fields are included only as supplementary references for sensor selection, model calibration, and inspection methods. This article synthesizes diagnostic indicators for nitrogen, phosphorus, and potassium, together with their associated physiological responses and spectral characteristics; compares the performance of hyperspectral, multispectral, and machine learning methods at the leaf, plant, and canopy scales; and examines fixed measurement, stop-and-go mobile inspection, continuous motion imaging, and autonomous plant revisitation. Existing studies have established a solid foundation for nutrient content retrieval, deficiency identification, and mobile monitoring. However, several challenges remain inadequately addressed under continuous inspection conditions, including radiometric–geometric joint calibration, plant identity preservation, acquisition of multi-element chemical truth values, model generalization across growth stages and greenhouse types, and long-term performance evaluation. Future work should refine standardized protocols for dynamic data collection and water–fertilizer environmental control, integrate mechanistic constraints with data driven approaches, and incorporate plant re-identification, spatiotemporal registration, uncertainty quantification, and online calibration. These efforts will contribute to constructing a long-term stable and comparable nutritional diagnostic system, thereby advancing the transition of facility vegetable nutritional monitoring from single-time static measurements toward continuous, traceable, and autonomously patrolled systems that may ultimately support precision irrigation and fertilization management after appropriate independent validation. Full article

Artificial intelligence is becoming deeply embedded in higher education, yet how the characteristics of AI-based e-learning systems relate to students’ perceived learning effectiveness and satisfaction remains insufficiently understood. This study examines the relationships of AI Functionality Compatibility, AI Instructional Process Coverage, and AI-Assisted Learning Cognitive Usability with Perceived Online Learning Effectiveness and Online Learning Satisfaction. Data from 384 students at Chinese universities were analyzed using a two-stage approach combining partial least squares structural equation modeling and artificial neural networks (PLS-SEM-ANN). The results showed that all three system characteristics were positively associated with perceived learning effectiveness, with instructional process coverage showing the strongest relationship. Cognitive usability also had a significant direct association with learning satisfaction, whereas functionality compatibility and instructional process coverage showed significant indirect effects through perceived learning effectiveness. The findings reveal an outcome-oriented pattern in which perceived learning effectiveness occupies a central position between system characteristics and satisfaction. This study extends understanding of AI-supported learning systems by emphasizing the alignment of technical functions with pedagogical processes and learners’ cognitive needs. It also provides practical guidance for universities and developers seeking to better align the design and evaluation of AI-based e-learning systems with learners’ instructional and cognitive needs. Full article

Advanced battery management systems (BMS) demand high-precision online prediction of lithium-ion battery state of health (SOH). To address engineering challenges where deep learning and multi-physics models struggle with edge deployment on automotive-grade microcontrollers (MCUs) due to high memory footprints and inference latency, and where traditional methods obscure true capacity regeneration during rest periods, this study proposes a lightweight, low-latency, and high-precision Adaptive Closed-loop Correction Model (ACCM). First, the equal-voltage-drop discharge time (EVDDT) and maximum temperature rise (ΔT) are extracted as indirect health indicators (HIs). Subsequently, a double-exponential model with explicit capacity regeneration compensation (CRC-DEM) is established, utilizing a Hybrid Particle Swarm Optimization-Nelder-Mead (Hybrid PSO-NM) algorithm for initial parameter identification. Furthermore, through an event-triggered adaptive closed-loop correction driven by HIs, an asymmetric time-window weighted objective function is employed to successfully overcome the historical inertia trap typical of the Open-loop Extrapolation Prediction Model (OEPM). Blind testing on NASA battery aging datasets demonstrates that this closed-loop model accurately tracks capacity regeneration while restricting the maximum root mean square error (RMSE) for SOH and capacity within 1.4% (average RMSE: 1.1%). With a peak dynamic memory of merely 4.08 KB and a maximum online inference latency of 0.1045 ms, the proposed model demonstrates superior lightweight efficiency and competitive accuracy compared to traditional deep learning approaches (e.g., CNN, BiLSTM, CNN-LSTM-Attention). Full article

Accurate state of charge (SoC) estimation is of primary importance in terms of safe and efficient management of energy storage systems (ESSs). In this regard, data-driven frameworks offer the advantage of rapid execution during online operations. Nevertheless, their deployment on resource-constrained embedded systems is often hindered by the strict memory and processing limitations of low-cost hardware. This article proposes a three-stage pipelined SoC estimation framework incorporating a reinforcement learning (RL) primary stage, least squares boosting (LSB) secondary residual corrector, and ultimate linear three-point interpolation (3p-InT) stage. The RL phase utilizes a twin deep delayed deterministic policy gradient neural network (TD3NN)-based agent along with a customized reward function. The inference part of the framework is deployed on two separate embedded platforms, i.e., an STM32F411RE microcontroller (MCU) and Digilent Nexys A7-100T FPGA through automatic C code and hardware description language (HDL) code generation features in MATLAB/Simulink, respectively. The efficacy of the proposed framework is evaluated using a Panasonic 18650 lithium-ion battery (LiB) and a battery-powered drill load profile (BPD-LP). Across the four hardware scenarios, the accuracy of the proposed framework is preserved, with the maximum %RMSE deviation not exceeding 0.08 percentage points. The RMSE value remains within 1.80–1.82% for the LiB dataset and within 0.76–0.84% for the BPD-LP, irrespective of the platform or the arithmetic format. As for the resource footprint, the fixed-point implementation more than halves the FPGA logic with respect to the floating point (27.33% against 65.55% of the LUTs), at the cost of a comparatively higher DSP usage (15% against 7.08%). On the MCU, it trades additional flash memory (31.51% against 25.23%) for a 2.6-fold smaller RAM footprint. The framework’s reward function, hyperparameters, and architecture are kept unchanged across both datasets, indicating that the same configuration can be generalized across both profiles without requiring dataset-specific re-tuning. Moreover, the detailed hardware deployment findings provide a practical insight into the hardware and arithmetic format selection for an accurate embedded SoC estimation framework. Full article

Attention-deficit/hyperactivity disorder (ADHD) involves difficulties in sustaining attention and resisting distraction. This has motivated the development of feedback-driven environments for cognitive control training. Integrating electroencephalography (EEG) sensors into Virtual Reality (VR) serious games for cognitive therapy remains relatively underexplored and requires reliable, non-invasive brain–computer interfaces. The existing solutions use multi-channel systems that primarily suffer from requiring complex hardware, while not combining motor imagery (MI) with concentration levels. Therefore, this case study evaluates the feasibility and data quality of a lightweight, cost-effective sensor configuration for real-time control of mental state. A non-invasive, eight-channel OpenBCI Cyton board was integrated with an EEG cap using the international 10–20 placement system, alongside a Meta Quest 2 headset, to capture MI and concentration signals directly from the user’s scalp. Signal acquisition was hindered by high impedance and channel railing, which required conductive gel mitigation, while mechanical tension from the VR headset strap introduced motion artifacts and noise. Nevertheless, under stable signal conditions, the optimized eight-channel sensor setup achieved a subject-specific online classification accuracy of up to 90% using the deep learning model “EEGNet”. The findings demonstrate the technical feasibility of acquiring and classifying EEG activity using a low-cost eight-channel sensor configuration in an interactive VR-BCI Serious Gaming application, provided that skin–electrode impedance and mechanical sensor interferences are managed. The results provide a basis for future investigation of such systems in cognitive-training applications, while further studies, including clinical evaluations, are required to assess their applicability in therapeutic contexts. Full article