In this study, we simulated slug flow using interFoam in OpenFOAM, which enables 3D twophase flow analysis, and examined flow characteristics such as velocity and turbulent kinetic energy (TKE). Numerical simulations were conducted under a total of 12 superficial Reynolds number conditions. The simulations reproduced the phenomenon in which the elongated bubble region behind the liquid slug formed in a narrower and longer shape than that at the front, confirming that qualitative reproduction was feasible. As ReSL increased, slug length decreased and the number of slugs within the pipe increased. Under the Re-SG = 1377 condition, the velocity ratio U/Um inside the liquid slug was below 1.2, whereas U/Um >= 1.5-2.0 appeared in the elongated bubble region. Under Re-SG = 4320, the elongated bubble region exhibited U/Um >= 2.0, while the liquid slug showed a relatively low velocity ratio of approximately U/Um approximate to 1.0. TKE/U-d(2) was highest primarily along the liquid-gas interface. Under the Re-SG = 1377 condition, the liquid slug interior showed low TKE/U-d(2) < 0.05, whereas values exceeding 2 occurred at the liquid-gas interface. For Re-SG = 4320, TKE/U-d(2) = 0.15-0.3 or higher appeared at the elongated bubble region, while the liquid slug interior showed TKE/U-d(2) approximate to 0.01-0.1. For V-L/V, under Re-SG = 1377, the value increased with ReSL, and the temporal fluctuation between the maximum and minimum values was large. Under ReSG = 4320, V-L/V also increased with ReSL, but the fluctuation was not pronounced. These results indicate that a three-dimensional numerical model for two-phase flow can reasonably reproduce the complex behavior and flow characteristics of slug flow. From these findings, future work involving the correlation analysis of numerical data and drone-based imaging data may contribute to identifying crack formation mechanisms in underground stormwater tunnels.
By combining cultural symbolism with practical value, the marine character intellectual property (MCIP) created from regional maritime resources is increasingly recognized as a strategic asset for coastal communities. In this study, we investigated three representative South Korean cases-Dokdo Guard Kangchi in Ulleung County, HOBOT in Yeongdeok County, and Padossi in Pohang City-to understand their contributions to regional identity, community engagement, and sustainable development. A qualitative approach was adopted, integrating a literature review and multicriteria case studies, with selection criteria emphasizing cultural and educational values, creativity and innovation, community engagement, and long-term viability. MCIP acts as symbolic capital, transforming local heritage, ecological values, and community narratives into accessible forms through media, merchandise, and public initiatives; it enhances environmental literacy (particularly among children and adolescents) by embedding ecological themes in culturally resonant storytelling. Simultaneously, it stimulates the local economy by attracting tourists, generating licensing opportunities, and strengthening regional brand recognition. Such IP fosters cultural resilience, promotes ecological stewardship, and creates scalable opportunities for global dissemination. In sum, MCIP represents an effective model for integrating cultural heritage into the contemporary content industry, offering coastal communities a sustainable pathway to preserve their identities while engaging with both domestic and international audiences.
The integration of AI into automated optical inspection enables the enhancement of printed circuit board manufacturing efficiency, accuracy, and cost-effectiveness. We developed an AI-assisted, vision-based defect sensing system that conducts multiclass electronic component recognition and surface defect detection by coupling an industrial optical sensing module with an enhanced You Only Look Once version 9-e (YOLOv9-e) deep learning architecture. The developed system enables the integration of controlled ring lighting and a high-resolution CMOS visual sensor to overcome image degradation in raw sensory signals, establishing a highly accurate, noncontact optical inspection concept. Printed circuit board samples containing diverse soldering defects from a national technical examination framework were utilized to compile a dataset of 220 images. Comparative YOLOv9-e outperformed YOLOv7, achieving a component recognition mean average precision at the intersection over union threshold of 0.50 (mAP@0.5) of 93.6%, an F1-score of 90.0%, and a defect detection accuracy of 89.8%. Although the noncontact sensing configuration of the developed system provides robust, real-time diagnostic capabilities, limitations exist, including dataset diversity and susceptibility to ambient illumination variations during sub-millimeter solder wetting inspection. To address the limitations, computational structural re-parameterization in the network layers is required to preserve critical geometric reflections.
We developed a decentralized, intelligent home appliance maintenance system driven by federated learning (FL) and multimodal sensor fusion. While modern IoT-enabled appliances generate vast amounts of high-fidelity data, their utilization is often hindered by privacy regulations and the proprietary nature of manufacturer-specific diagnostic information. To overcome these challenges, we developed a federated feature disentanglement system that isolates universal fault patterns from manufacturer-specific signatures, enabling secure cross-brand knowledge sharing. The system integrates time-series data from embedded physical sensors, including three-axis accelerometers for vibration and NTC thermistors for thermal profiling, with visual fault images and textual repair logs. The system showed a diagnostic accuracy of 94.5% and a predictive maintenance accuracy of 88.2%, outperforming centralized and vanilla FL baselines. The system also exhibits high efficiency in knowledge transfer, requiring only 4500 samples for a new manufacturer to reach stable performance with a 75% reduction in data requirements compared with traditional methods. With a privacy protection value of 3.1 and a subsecond system response time of 0.85 s, the system serves as a foundation for the development of next-generation privacy-aware, self-describing smart sensors that can deliver real-time, cross-platform appliance health management. Despite these results, the study is limited by the assumption of stable network connectivity among edge nodes. Therefore, it is necessary to optimize the system for intermittent connectivity and reduce the computational load for lower-tier sensor hardware.
MedMNIST, which is a collection of medical classification datasets, contains medical images obtained with a variety of techniques, such as computed tomography and optical coherence tomography, and images captured using a microscope. Compressing MedMNIST datasets using a dataset distillation method is a challenging task in the images per class (IPC) = 1 setting. Regarding this problem, methods that do not use gradients to update a compressed dataset largely remain unexplored, especially in MedMNIST. We propose a simple dataset distillation method based on the Deep Learning Evolutionary Algorithm (DL-EA) to optimize synthetic images mainly focused on MedMNIST. Our proposed method updates synthetic images and evaluates them using a subset of training data. Since DL-EA is a neural network (NN) training method, it cannot be applied to dataset distillation directly so we extended it by updating synthetic images rather than NN weights; this is our main contribution. To extend our optimize synthetic images. (2) To reduce the computational cost of each generation, we used only the subset of the training dataset. Our method is evaluated on MNIST and seven medical datasets included in MedMNIST. These datasets have been evaluated by gradient-based dataset distillation. We mainly focus on IPC = 1. Our method has a higher accuracy for MNIST and seven medical datasets included in MedMNIST than does strong random selection.
Rapid population aging and industrial restructuring make coordination between public elderly care services (PECS) and the elderly care industry (ECI) critical for improving healthcare outcomes in transition regions. In this study, a modified quadruple-subject coordination model (government-market-family-unit) was introduced to develop a sensor-integrated long short-term memory (LSTM)-deep neural network (DNN) model and evaluate the coupling coordination degree (CCD) of elderly healthcare systems. Multi-modal sensor data, including the global navigation satellite system (GNSS) parsing of National Marine Electronics Association (NMEA) 0183 Standard, the 2.4 GHz active RF identification tracking of electronic product code strings, mattress-embedded piezoresistive force sensors, dual-beam passive infrared (PIR) arrays, and 13.56 MHz near-field communication (NFC) handshakes, was collected across 14 prefecture-level cities in Liaoning Province, China. To fuse asynchronous data, a two-stage synchronization protocol employing rolling median filtering and bucket-aggregation grids was implemented before conducting sequence learning. Liaoning's CCD increased from 0.54 in 2018 (barely coordinated) to 0.79 in 2023 (moderately coordinated), progressing through preliminary, policy-driven, and optimization stages. The hybrid model eliminated gradient instability, reducing the root mean square error by 29.3% and achieving an R-2 of 0.923 compared with models lacking physical sensing layers. Governance constraints were identified, including urban-rural resource imbalance and a 1.2-year policy execution latency. By integrating explicit time-delay variables and Kalman filtering and adaptive drift correction, the developed model provides a reproducible, scalable approach for sensor-based healthcare governance evaluation in industrial transition regions.
Rotator cuff tear (RCT) is a common cause of shoulder pain resulting from tendon or muscle injury. Although magnetic resonance imaging (MRI) is the clinical gold standard for RCT diagnosis, its high cost and limited accessibility often delay treatment. While conventional X-ray imaging sensors are widely available and cost-effective, they lack direct visualization of soft tissues. To maximize the diagnostic utility of radiographic sensor data, we propose a two-stage deep learning framework that leverages indirect indicators, particularly greater tuberosity sclerosis (GTS), to assess RCT severity. In the segmentation stage, a multi-attention-gated U-Net with full-scale skip connections accurately delineates GTS regions from the X-ray sensor images. In the classification stage, a dual-branch convolutional network integrates the acquired sensor data and GTS masks using spatial and channel attention mechanisms to classify RCTs into partial-or full-thickness tears. The segmentation model achieved a Dice coefficient of 0.835 and an accuracy of 0.998, outperforming several state-of-the-art methods. The proposed classification network reached an overall accuracy of 0.941, which surpasses those of previously reported MRI-based and X-ray-based approaches. This framework demonstrates how advanced computational technologies can significantly augment the diagnostic capabilities of standard X-ray sensing systems, enabling accurate and efficient RCT assessment, reducing reliance on MRI, and supporting timely clinical decision-making.
In this study, we quantitatively evaluated the applicability of an integrated operational framework combining a heavy-class remotely operated vehicle (ROV) with manned diving systems, including the surface-supplied diving system (SSDS) and self-contained underwater breathing apparatus (SCUBA), under predicted tidal conditions at the Sewol ferry disaster site. A 31-day tidal-current dataset was analyzed to assess operational durations, phase-based deployment strategies, and personnel requirements. The results showed that integrated ROV-diver operations substantially expanded operational availability compared with standalone diving, increasing deployable time by 26.4% for SSDS and 54.7% for SCUBA. During neap tides, SSDS enabled long-duration deployment, whereas SCUBA maintained operational efficiency through rapid preparation and high mobility. Under spring tide conditions, the limited diving window made the heavy-class ROV the primary platform for search, inspection, and hazard-removal tasks. In this phase-based framework, sensor-derived information from the ROV supports underwater environmental assessment, target detection, hazard identification, and decision-making on the transition between ROV operations and diver deployment. Personnel analysis further showed that actual requirements were substantially lower than arithmetic estimates and represented only 6.6-10.1% of the Sewol response workforce, indicating the feasibility of coordinated ROV-diver missions. These findings provide a quantitative basis for developing standardized procedures and integrated command structures for large-scale underwater search and rescue operations.
Airborne bathymetric Light Detection and Ranging (LiDAR) systems have attracted attention as efficient surveying tools that acquire high-resolution and high-precision coastal topographic data more cost-effectively than traditional shipborne acoustic sounding or field surveys. Since laser pulses are refracted at the air-water interface, for the accurate registration of seafloor points, it is crucial to distinguish whether each return signal is from land or water at the waveform stage, before generating the point cloud. Conventional land-water discrimination techniques often rely on near-infrared (NIR) channel data for water-surface detection or water-body identification. However, NIR signal reliability is often compromised by specular reflection from water surfaces, and many recently developed sensors employ only a single green laser wavelength owing to system miniaturization and weight reduction. This situation underscores the necessity for land-water discrimination techniques that use only single green-channel waveform information. In this study, we analyzed various waveform features extracted from individual waveforms acquired with the Seahawk airborne bathymetric LiDAR system across coastal areas with varying water depths and turbidities. These waveforms were decomposed into Gaussian components, from which features were extracted and used in machine learning classifiers to evaluate their versatility and effectiveness for land-water discrimination under diverse coastal conditions. Four tree-based machine learning models-decision tree, random forest, XGBoost, and LightGBM-were evaluated using a stratified cross-validation scheme for performance assessment. All models achieved a high validation accuracy of approximately 0.99, demonstrating discriminative capability based on waveform features. In comparative evaluations considering both test accuracy and computational efficiency, LightGBM showed the most balanced performance, indicating its suitability as a general-purpose model for waveform-based land-water discrimination.
In this study, we introduce the Graphiti Model Context Protocol (MCP) Server, an AI-assisted programming framework that integrates a time-aware knowledge graph with the MCP to enhance contextual sensing and semantic reasoning in intelligent code generation. The system addresses key limitations of existing AI agents in long-term software engineering tasks- particularly the lack of persistent memory and insufficient contextual perception. The proposed architecture consists of an MCP client, the Graphiti MCP core server, a Neo4j-based knowledge graph database, and semantic sensing modules connected to the OpenAI Application Programming Interface. Together, these components enable dynamic knowledge retrieval, cross-file context fusion, and adaptive code assistance. Experiments conducted on 1000 open-source software projects demonstrate statistically significant improvements: code completion accuracy (CCA@1) increased from 45.3 to 68.5%, contextual relevance score improved from 3.0 to 4.2, and task completion time decreased by approximately 35.5%. Furthermore, a mathematical model is developed to describe how graph-based knowledge retrieval enhances effective memory sensing and contextual stability, providing a theoretical foundation for intelligent programming. The results verify that the Graphiti MCP Server offers substantial potential for advancing context-aware and sensor-like AI systems with semantic sensing and adaptive contextual reasoning capabilities in software development environments. However, current AI-assisted programming systems still suffer from several limitations, including restricted longterm memory, fragmented contextual understanding, and insufficient semantic perception in large-scale software development environments. To address these challenges, we propose a knowledge-graph-based context-aware semantic sensing and contextual reasoning framework that integrates semantic retrieval, persistent memory, and adaptive context fusion through the Graphiti MCP Server architecture. Although the proposed framework significantly improves contextual awareness and programming efficiency, challenges related to large-scale knowledge graph maintenance, multimodal information integration, and reasoning scalability remain important directions for future research.
In intelligent human environments, image sensors are frequently hindered by occlusions, such as masks, which degrade the reliability of facial data for security and interactive support systems. In this paper, we propose a unified framework designed as a robust support system that integrates an occlusion segmentation network with a face image restoration network. To facilitate deployment in resource-constrained sensing nodes, the segmentation network employs depthwise separable convolutions to ensure computational efficiency while leveraging residual connections for multi-scale feature fusion. On the basis of precisely localized occluded areas, a generative adversarial network is introduced to reconstruct facial structures with high fidelity. The generator incorporates two novel feature enhancement components: a hybrid attention aggregation module, which strengthens global semantic consistency within skip connections, and a multi-scale spatial attention module, designed to capture fine-grained textures from sensor data across different spatial scales. Experimental results on the CelebFaces Attributes High-Quality (CelebA-HQ) dataset demonstrate that the proposed system effectively restores masked facial regions, achieving a PSNR of 35.01 dB and an SSIM of 0.931 under challenging 35-45% occlusion ratios. By significantly enhancing visual fidelity and recognition robustness, this framework provides a reliable solution for real-world vision-based support systems in human-centric environments.
Remote monitoring in lifecare presents a vision of human behavior interactions in complex situations that were previously overlooked. In this research, we designed lifecare algorithms that recognize the activities from processed video sequence images. This groundbreaking technology enables remote monitoring systems to collect reliable data that can be applied across fields such as healthcare, sports, and security. We integrate an embedded hidden Markov model (E-HMM) with visual-sensor-based data input to enhance the efficiency of human behavior interaction systems. The proposed method begins by performing a hue saturation value color transformation to improve the clarity of video frames. The silhouette is extracted using hybrid techniques with sensor data, and signal features are extracted using texton maps, a local intensity order pattern, and oriented features from accelerated segment test and rotated binary robust independent elementary features. Fuzzy optimization is then carried out to choose the most discriminative signal features. An E-HMM is trained to identify actions correctly according to the given functions. Furthermore, since the suggested approach monitors the order of actions, it uses time-related data, which results in improved detection results, even in the presence of occlusions or when actions are performed at various speeds and scales. The sensor data used in the experiment, combined with the recognition algorithms, achieved the following results: Shakefive2: 0.97%, HMDB51: 0.90%, and Okutama Action: 0.68, 0.94, and 0.82%.
Running biomechanics was investigated using sensor technology to clarify the relationships between important parameters and injury prevention. Through secondary data analysis and case evaluations, validated thresholds of biomechanical variables were established: ground reaction force (GRF) loading rate (<= 65 body weight/s), gait asymmetry index (<= 15%), and ground contact time (<= 250 ms). These thresholds can be used as indicators of running efficiency and injury risk. Sensor data demonstrated negligible measurement error and offered reliable biomechanical information, confirming their suitability for real-time monitoring and intervention. GRF and loading rates were identified as essential predictors of injury susceptibility. Furthermore, machine learning models trained on sensor data accurately detected biomechanical abnormalities, supporting the integration of automated monitoring systems into injury prevention strategies. The sensor-based approach enables evidence-based guidelines for parameter interpretation, advances methodological validation, and promotes standardized mathematical modeling. It also facilitates intelligent monitoring programs, individualized training, and personalized prevention protocols. The effective integration of machine learning with wearable sensors requires devices capable of delivering real-time, personalized feedback on cadence, ground contact time, and related metrics. To establish validated thresholds, three evaluations were conducted: the analysis of the Gutenberg Gait Database, the assessment of the Human Activity Recognition-3 dataset, and the systematic meta-analytical synthesis of 156 peer-reviewed studies published between 2015 and 2025.
