ETRI Journal最新文献

The growing adoption of edge AI in smart city applications such as traffic management, surveillance, and environmental monitoring necessitates efficient computational strategies to satisfy the requirements for low latency and high accuracy. This study investigated GPU sharing techniques to improve resource utilization and throughput when running multiple AI applications simultaneously on edge devices. Using the NVIDIA Jetson AGX Orin platform and object detection workloads with the YOLOv8 model, we explored the performance tradeoffs of the threading and multiprocessing approaches. Our findings reveal distinct advantages and limitations. Threading minimizes memory usage by sharing CUDA contexts, whereas multiprocessing achieves higher GPU utilization and shorter inference times by leveraging independent CUDA contexts. However, scalability challenges arise from resource contention and synchronization overheads. This study provides insights into optimizing GPU sharing for edge AI applications, highlighting key tradeoffs and opportunities for enhancing performance in resource-constrained environments.

This study investigates intelligent reflecting surface (IRS)-assisted cognitive radio (CR) short-packet communication (SPC) networks. In the secondary network, a base station communicates with a user with support from an IRS in a Nakagami-$�m$ fading environment. We aimed to evaluate the block error rate (BLER) performance of the secondary network under a limited interference temperature for scenarios with the presence and absence of a direct base station-user link. To this end, we employed two approaches: (i) conventional mathematical analysis and (ii) data-driven performance evaluation. For the former, closed-form expressions of the average BLER and asymptotic average BLER of the secondary user were derived analytically. For the latter, a deep neural network (DNN) model was constructed to evaluate BLER as a regression problem. A Monte Carlo simulation approach was adopted to verify the accuracy of the derived analytical BLER and DNN-based BLER performance evaluations. We determined that coherently utilizing both direct and IRS-reflecting links can significantly enhance the BLER performance of IRS-aided CR SPC networks.

When using lazy learners based on the Mahalanobis distance (MD) function for process fault detection (FD), due to the curse of dimensionality, type I errors can increase significantly as the number of process variables increases. In high-dimensional data spaces, certain regions exist in which data samples are sparsely distributed. From the perspective of dense regions, the outlierness (i.e., degree of being statistical outliers) of samples in sparse regions increases as the data dimensions increase, leading to unstable estimations of classical covariance matrices for calculating MD function values. To solve this problem, a lazy learning method is proposed based on a robust MD function, where robust covariance matrices are estimated using a minimum covariance determinant method. Here, k-nearest neighbors and local outlier factor are employed as baseline learners. The proposed method can be applied to all types of lazy learning techniques. To verify FD performance, the proposed method is applied to two benchmark processes. The experimental results show that the proposed method can perform FD on very high-dimensional processes successfully without rapid increases in type I errors.

Sewer infrastructure management is essential for public health, environmental protection, and urban stability. Aging networks and the impacts of climate change emphasize the need for advanced management solutions. Traditional methods, such as periodic inspections and reactive maintenance, are insufficient to address the complexities of modern sewer systems. This study surveys intelligent-sensor-based management technologies aimed at improving sewer infrastructure. Key technologies include Internet-of-Things-driven data collection, machine learning and deep learning analytics, cloud and edge computing, and autonomous robotics. Based on case studies from South Korea, Germany, Japan, and the United States, the practical benefits of these technologies were explored, including real-time monitoring and predictive maintenance, as well as challenges such as sensor durability, robotic mobility, and data analysis limitations. Rather than proposing solutions, this study evaluates the current state of these technologies and identifies gaps that require further research and innovation. It provides a comprehensive overview that serves as a valuable resource for researchers and practitioners and contributes to the advancement of sustainable and efficient sewer management systems.

User reading status provides valuable insights into cognitive processes. Most reading classification methods rely on electrooculography (EOG). However, classifying EOG signals in uncontrolled environments poses challenges because of noise and limited data. To address this issue, methods based on self-supervised learning or nested architectures have been proposed. However, their performance is often limited because they do not optimize the model in an end-to-end manner. Therefore, we propose an end-to-end network for two-dimensional (2D) signal images generated from reshaped EOG signals. A 2D signal image was generated by reshaping EOG signals to incorporate both horizontal and vertical eye movements along with their magnitudes. We designed a ResNet-based network to classify 2D signal images and introduced data augmentation techniques commonly used in image classification tasks. Our experiments, conducted using a publicly available dataset, evaluated various factors such as network structures, segmentation strategies, sampling rates, and sensor modalities. The results demonstrate that the proposed approach significantly improves classification accuracy compared with existing methods.

Occupancy detection systems are crucial for optimizing energy efficiency in smart cities and buildings but often face privacy and data dependency challenges. YOLO (you only look once), a widely used real-time detection framework, relies on identifiable image data and labeled datasets. This study proposes a privacy-preserving, labeling-free occupancy sensor using a time-of-flight (ToF) camera, and a clustering algorithm. Positioned above doorways, the ToF camera captures depth data that inherently protect privacy by avoiding identifiable information. Using the mean shift clustering algorithm, it performs real-time detection and tracking without labeled data, generating bounding boxes for movement analysis. Unlike traditional ToF-based or unsupervised methods, the proposed system adapts dynamically to varying occupant behaviors and environmental conditions for robust real-time detection. Experimental results show that the proposed method achieves over 90% accuracy in standard single-entry and exit scenarios. By addressing existing limitations, it offers a data-efficient, privacy-sensitive solution for building digital twins in energy optimization and resource management.

In skeleton-based action recognition, most state-of-the-art models are based on graph convolutional networks (GCNs), and propose various graph topologies to capture the inherent relationships between joints more effectively. However, existing GCN-based methods are constrained by their reliance on independently proposed graph topologies, and often overlook the potential benefits of incorporating various graphs. To address this limitation, we propose a dynamic partitioning GCN (DPGCN) that can capture the dynamic dependencies of the skeletal structure and learn the relationships among subgraphs. The DPGCN, inspired by dynamic convolution, includes a dynamic partitioning graph convolution (DP-GC) designed to extract features from subgraphs and predict dynamically adjusted weights. DP-GC assigns predicted weights to multiple kernels and combines them, allowing the model to focus on the structural patterns. Our proposed DPGCN outperforms or achieves a performance comparable to state-of-the-art methods on three benchmark datasets: NTU RGB+D, NTU RGB + D 120, and NW-UCLA.

Wireless power transfer (WPT) technology offers a promising solution for powering electronic devices without a physical connection. However, achieving high power-transfer efficiency (PTE) while minimizing electromagnetic interference (EMI) remains a critical challenge, especially for flexible and unrestricted device positioning. This study explores the use of coil rotation and phase-shift control to optimize the PTE by adjusting the transmitter (TX) coil orientation and phase shifts. Analytical expressions based on the Neumann formula are employed to derive the mutual inductance between two coaxially aligned coils with varying receiver (RX) coil orientations. A prototype magnetic resonance WPT (MR-WPT) system is developed to validate the feasibility of the proposed efficiency enhancement methods. The simulation and experimental results demonstrate that optimizing the TX coil phase-shift and coil-rotation angle can maximize the RX voltage and improve the PTE by approximately 30%, while also reducing EMI levels.

Under ubiquitous smart environments, the convergence of mobile ad hoc networks (MANET) and infrastructure networks enables new communication patterns. In this hybrid MANET (H-MANET) platform, gateways critically affect network performance. We address the gateway selection problem by proposing a novel decision mechanism that considers multiple metrics. Using a multi-criteria decision method and bargaining game theory, we develop a novel gateway selection algorithm. First, routing paths are discovered. Second, decision criteria—route distance, queue length, connectivity degree, and link complexity—are evaluated. Third, each gateway's adaptability is assessed through the combination of Kalai–Smorodinsky and Nash bargaining solutions. Finally, the most adaptable gateway is selected for data transmission. Our main contribution is integrating both bargaining solutions' concepts for multi-criteria-based gateway selection. Simulation results demonstrate the performance benefits of our proposed approach over existing methods. The proposed method can also address other real-world multi-criteria decision problems.

Applications in various fields, such as deep learning and scientific computing, naturally exhibit data access patterns along both the row and column dimensions of static random access memory (SRAM). However, traditional SRAM architectures only support asymmetric access, typically in the row dimension. Accordingly, to perform operations that require both row-wise and column-wise data, SRAM must activate multiple rows sequentially to obtain column-wise data. Such time-consuming operations significantly degrade not only the performance but also the energy efficiency of graphics processing units (GPUs). In this study, we exploited monolithic 3D (M3D) integration to construct a large-scale SRAM architecture supporting accessing data in both the row and column dimensions (that is, multi-dimensional access [MDA]). When our M3D-MDA memory is utilized for GPU scratchpad memory, it provides average performance improvements of 68% and 23.4% for fundamental operations and workloads that exhibit MDA access patterns, respectively, compared with traditional 2D memory. The M3D-MDA memory also substantially reduces energy consumption.