NTNs are expected to become a fundamental component of next-generation communication infrastructures. Integrating NTN within the O-RAN framework provides notable benefits such as reduced payload complexity and lower development, launch, and replacement costs. Despite these advantages, significant challenges persist in extending the terrestrial O-RAN architecture into three dimensions. This paper examines these challenges and proposes a three-dimensional O-RAN architecture for future NTNs.
Accurate channel state information (CSI) at the base station (BS) is crucial for achieving high beamforming gains in massive multiple-input multiple-output (MIMO) systems. In frequency division duplex (FDD) systems, the BS relies on feedback from the users (UEs) to obtain CSI. However, the aggressive CSI compression and quantization at the UE limits the achievement of the optimal beamforming gains. To resolve this problem, we propose an autoencoder network that learns a compact, discrete representation of the channel for variable feedback rates. We train and evaluate our variable-rate autoencoder using a large synthetic dataset, generated through advanced 3D modeling tools and ray-tracing simulators. Simulation results demonstrate that the proposed network achieves more accurate channel reconstruction compared to competitive methods.
This paper presents a multi-sensor simulator for UAV communication that generates datasets incorporating UAV dynamics, sensor disturbances, and non-stationary processes. Leveraging these datasets, an extended Kalman filter (EKF)-based localization scheme is employed to address the nonlinear observation model. Experimental results demonstrate that the EKF significantly improves both velocity and angular estimation accuracy compared to noisy measurement.
The advancement of smart factories and multi-stage logistics has accelerated the need for efficient and adaptive coordination of autonomous mobile robots (AMRs). While most existing research has focused on homogeneous automated guided vehicles (AGVs) systems, the control of heterogeneous AMRs remains underexplored. This paper proposes a multi-agent reinforcement learning (MARL)-based control framework tailored for heterogeneous AMRs in smart factory environments. The framework incorporates AMR-specific characteristics such as field of view (FOV) into the observation space and models decision-making using a multi-discrete action structure with action masking to ensure feasibility. The reward function promotes efficient task execution by encouraging successful pickup, delivery, and goal-directed movement. Simulation results demonstrate that the proposed approach achieves stable learning, improves delivery completion rates, and reduces task execution time, validating its effectiveness in heterogeneous and complex factory settings.
Unmanned Aerial Vehicle (UAV) swarms are increasingly deployed in Industrial Internet of Things (IIoT) applications, but remain vulnerable to Positioning, Navigation, and Timing (PNT) spoofing and federated model poisoning, which threaten safety and mission reliability. Traditional cloud-centric or heavyweight blockchain solutions suffer from latency and scalability limits in UAV-class hardware. This paper proposes a hybrid framework that integrates Horizontal Federated Learning (HFL) with a lightweight DAG-based blockchain. The HFL layer enables privacy-preserving anomaly detection with Byzantinerobust aggregation, while the DAG blockchain ensures low-latency consensus, tamper-proof anomaly logging, and update reputation tracking. Experimental validation on a GNSS spoofing dataset shows 81.5% detection accuracy under non-IID data (vs. 52.5% FedAvg), 62% lower testing loss, and sub-second transaction latency (0.85s) with 25 MB lower communication cost per round. The proposed design establishes a unified, auditable, and scalable defense pipeline, advancing resilient UAV swarm communication infrastructures for IIoT logistics, inspection, and surveillance.
Monocular distance estimation enables 3D understanding from single-camera images, supporting navigation and safety applications. We present two RetinaNet-based single-stage models: DistinaNet and DistinaNet-BBoxExt, enhancing the RetinaNet architecture with distance estimation capabilities via ExtraHead and BBoxExtension strategies. Benchmarked on the KITTI dataset, DistinaNet outperforms state-of-the-art object-specific methods by 44.82% in MAE. Ablation studies show that adding the distance estimation task does not degrade detection performance. We also analyze FPN layer contributions, head architectures, and loss functions. Both models demonstrate strong generalization to unseen real-world driving scenarios, validated using data captured from a custom data collection pipeline. Code is available at https://github.com/jonher16/distinanetforreproduction.
This paper proposes a dynamic subgrouping method to overcome the limitations of conventional dual subgrouping approaches that combine Low Power Wake-Up Signal (LP-WUS) and Physical-layer Early Indication (PEI) for ultra-low power operation of IoT devices in wireless environments. Traditional fixed-K-value subgrouping fails to account for dynamic network characteristics such as device behavior patterns, mobility, and time-varying traffic load, leading to increased paging false alarms and reduced power efficiency. To address this, we present a method that dynamically configures PEI subgroups based on device features such as traffic frequency, DRX cycles, mobility patterns, and service types, while also adjusting the size and distribution of LP-WUS subgroups in real-time according to traffic load and device density within a cell. Furthermore, by calculating the K value dynamically based on real-time network status, the proposed approach minimizes inter-group interference and enables flexible group management. This strategy reduces paging false alarms and unnecessary wake-up signals, effectively lowering the power consumption of IoT devices while enhancing the overall efficiency of wireless networks.
In this paper, we address the routing problem in Low Earth Orbit (LEO) satellite networks under stringent feeder link (FL) constraints. LEO networks rely on satellites interconnected via inter-satellite links (ISLs) and connected to the ground through feeder links to a limited number of fixed-position ground gateways (GWs). While the large number of satellites provides spatial diversity for user access, the gateway infrastructure remains sparse due to economic and geographical limitations. As a result, FLs often become a bottleneck in the end-to-end (E2E) routing process, where traffic from user terminals (UEs) must traverse a satellite relay path before reaching a ground gateway. The dynamic nature of satellite positions, limited gateway visibility, and link congestion further complicate routing decisions. To address these challenges, we propose a routing strategy that explicitly accounts for FL availability during path construction. Unlike conventional methods that prioritize access or inter-satellite links first, our approach dynamically adapts routing decisions based on link feasibility, availability, and shared congestion.
Blood cells are one of the most important parts of our body to transport the nutritions that have been processed to the whole organs for our health. It is very important to do a precise identification of blood cells to keep an eye on patient conditions. In this research, I will identify the types of blood cells accurately using the combination of machine learning CNN (Convolutional Neural Network) object detection YOLO (You Only Look Once) from a dataset of blood cell images that called as BCCD (Blood Cell Dataset Count) and doing the classification method. There will be 3 kinds of results: red blood, white blood, also thrombocyte. This technique can help the doctor faster and easily identify blood cells for giving the needs of data diagnosis from the results of the bloods cells classification to detect various diseases.
As network traffic increasingly becomes encrypted, traditional traffic classification methods based on payload signatures are facing significant limitations. Although various approaches have been proposed to address this issue, they often encounter challenges such as suboptimal performance, extended processing time, and overfitting due to dataset-specific characteristics, making them impractical for real-world deployment. In particular, methods relying on the payload area for encrypted traffic classification are inherently constrained. To overcome these limitations, this paper proposes a classification method that leverages "Burst" characteristics—temporal patterns in encrypted traffic—to identify the application responsible for generating the traffic.
Wearable IoT devices enable continuous cardiovascular monitoring, but privacy risks, communication overhead, and data heterogeneity hinder AI-driven anomaly detection. This paper proposes a Federated Agentic Learning (FAL) framework that embeds autonomous AI agents into edge clients to enable adaptive participation and dynamic Differential Privacy (DP) management. Using the PhysioNet 2017 dataset with a CNN– LSTM model, FAL achieved an F1-score of 92.5%, surpassing standard federated learning (88.1%) and approaching centralized training (95.0%). Communication overhead was reduced by 42%, while fairness improved with 35% lower variance across clients. These results demonstrate that FAL is an efficient, robust, and privacy-preserving foundation for edge-based healthcare AI, paving the way for trustworthy cardiovascular monitoring in real-world deployments.
The adoption of artificial intelligence (AI) across various industries is rapidly expanding. However, data collection from manufacturing processes is often limited or costly, resulting in insufficient datasets for AI applications. To address this issue, data augmentation methods have been proposed. Recent advances in generative AI, including generative adversarial networks (GANs) and diffusion models, have emerged as promising alternatives. In this study, we investigate the feasibility of generative AI to expand image datasets. Specifically, we demonstrate the application of OpenAI’s image generation model to synthesize additional optical microscope images obtained during intermediate stages of semiconductor manufacturing. A comparative analysis and validation of the generated data were conducted to demonstrate the applicability of generative AI-based data augmentation. Our results indicate that generative AI provides an effective and practical approach for supplementing datasets in data-scarce environments.
This study investigates the optimization of annealing temperatures and subsequent fabrication steps to improve the ohmic contact resistance of Ti/Al/Ni/Au electrodes on AlGaN/GaN high-electron-mobility transistors (HEMTs). Ohmic contacts underwent rapid thermal annealing (RTA) at 880°C, 910°C, or 940°C for 30s. After device isolation process, further processing steps including 60-nm silicon nitride (SiNx) deposition, via etching, and metallization were conducted. Remarkably, these subsequent processes reduced the contact resistance by up to 40%, with an optimal final value of 0.44 Ω•mm achieved at the annealing temperature of 880°C. Additionally, a significant decrease in sheet resistance (~30%) was observed after SiNx passivation. This result underscores the critical role of integrated subsequent processes rather than isolated ohmic contact formation process in reducing final contact resistance in AlGaN/GaN-based HEMTs.
Next-generation 6G mobile communication systems will deliver high data rates and ultra-low latency services through sub-terahertz or terahertz bands. This makes developing wireless access protocols for 6G mobile communication systems a very challenging task. In this paper, we adopt Intel DPDK as a wireless access protocol development platform to implement a wireless access protocol with a processing speed of up to 50Gbps and present its test results. This work demonstrates the feasibility of implementing Tbps-class L2 protocols required for 6G networks using Intel Xeon CPUs and DPDK.
In sixth-generation (6G) networks, non-terrestrial networks (NTNs) using low Earth orbit (LEO) satellites are expected to be pivotal in extending coverage and service diversity. Most existing studies on LEO-based communications focus on downlink (DL) services within the Long-Term Evolution (LTE) framework. This paper addresses an NTN-terrestrial network (TN) coexistence scenario based on the fifth-generation (5G) orthogonal frequency division multiple access (OFDMA) architecture and proposes a hierarchical reinforcement learning-based resource allocation and scheduling scheme (HRL-RAS). The HRL-RAS optimizes resource distribution to meet heterogeneous user demands and improve key performance indicators (KPIs) such as data rate for future integrated NTN-TN systems.
In this paper, we study a spectral efficiency (SE) maximization problem in downlink aerial simultaneously transmitting and reflecting (STAR)-reconfigurable intelligent surface (RIS)-assisted massive multiple-input multiple-output (mMIMO)-rate-splitting multiple access (RSMA) networks. The formulated problem belongs to the non-convex mixed-integer class due to the maximum allowable portion of common SE, power budget limitation at the flying-base station (F-BS) and the discrete phase shift at the flying-STAR-RIS (F-SR), which is difficult to solve optimally. To tackle this problem, the formulated problem is relaxed and then decoupled into a phase-shift and a beamforming subproblem, and then solved alternately. A bisection search method is applied to address the phase shift sub-optimization problem, while the inner approximation (IA) technique is utilized to solve the beamforming sub-problem by transforming the original non-convex formulation into more tractable forms. Simulation results show that the proposed scheme can improve SE performance. Besides, the impact of the number of F-BS antennas and F-SR elements on the considered system is also evaluated.
We propose a vehicular traffic event transmission method for Vehicle-to-Infrastructure (V2I) local networks, using a fixed 20 bytes IPv4 option field to carry vehicular traffic event information such as car accidents, vehicular traffic congestion, road construction, and emergency situations. The motivation is to address critical limitations in current V2I communication systems, which suffer from high processing latency due to complex payload parsing, dependency on centralized servers that create bottlenecks, and delays in safety critical message transmission that can compromise road safety. In the proposed protocol, vehicles embed vehicular traffic event information directly into IPv4 option fields and transmit packets through roadside infrastructure. Intermediate nodes forward these packets based solely on header inspection without payload processing or server communication, enabling direct event transmission across the network. The protocol was implemented and tested in a laboratory environment, and its end-to-end operation was successfully verified.
This study explores the potential of applying industrial AI using small data-based mathematical models in data-constrained environments. Mathematical models for predicting slurry viscosity, particle size distribution, and color were developed based on actual process data. The proposed models achieved high prediction accuracy with limited data, supporting effective process monitoring and data-driven decision-making. These findings suggest that small data-oriented mathematical modeling is a viable approach for the practical deployment of industrial AI.
Traffic elements such as fluctuating vehicle density play a crucial role in shaping commute durations. High congestion levels can increase delays by 12–55%. Modifying traffic density across highway networks offers the potential to expand road capacity and ease travel time; however, this relationship is complicated by factors such as driver behavior, vehicle dynamics, and induced demand. As autonomous vehicles (AVs) gain prominence, understanding how traffic density affects their operation becomes essential for optimizing their benefits and informing the development of intelligent transportation infrastructure.This study employs reinforcement learning (RL)-based AV models to explore how traffic density influences key traffic metrics, including average speed, collision frequency, and lane discipline (e.g., adherence to the right-most lane). Experiments are conducted using two established RL algorithms—Deep Q-Networks (DQN) and Proximal Policy Optimization (PPO)—within the HighwayEnv simulation environment. Traffic density is varied systematically (with values = 1.0, 2.0, 3.0) along with a randomized Density-Mix condition, while maintaining a consistent four-lane highway setup.Findings indicate that lower densities generally lead to better traffic behavior, though the extent of improvement is influenced by the RL method used. PPO exhibits greater robustness and adaptability, especially in high and variable-density settings, whereas DQN shows strong performance in sparse traffic but struggles under congestion. Notably, training PPO under the Density-Mix scenario results in consistently balanced behavior across all traffic conditions. These results highlight the critical role of both algorithm selection and exposure to diverse environments in training resilient and effective autonomous driving systems.
Semiconductor wafer fabrication involves complex scheduling challenges with hundreds of processing steps, reentrant flows, and stringent operational constraints. Traditional dispatching rules lack adaptability to dynamic fab conditions, while optimization methods face computational scalability issues. This paper proposes a reinforcement learning framework that formulates wafer scheduling as a Markov Decision Process (MDP) with comprehensive state representation capturing equipment utilization, queue dynamics, and operational constraints. The multi-objective reward function balances throughput, cycle time, tardiness, and equipment utilization while respecting batching and setup requirements. Experimental evaluation demonstrates significant improvements over traditional methods, achieving 5.3% higher throughput, 21% reduction in tardiness, and superior equipment utilization across multiple performance indicators. The stable learning convergence validates the effectiveness of the proposed approach for dynamic semiconductor manufacturing environments.
