Taiwan Association of Engineering and Technology Innovation: E-Journals
Not a member yet
887 research outputs found
Sort by
Differentiated QoS Provisioning in Wireless Networks Based on Deep Reinforcement Learning
Get PDFWireless networks manage performance by adjusting the contention window, as they cannot directly detect collisions. Traditional contention window adjustment algorithms, such as the Binary Exponential Backoff (BEB) algorithm, may lead to lower throughput when multiple services with varying bandwidth demands coexist. To address this issue, this study aims to enhance network throughput by enabling differentiated bandwidth allocation for various services. Using deep reinforcement learning, the state space, action space, and reward functions are defined to optimize this differentiation. These definitions are integrated into the Deep Deterministic Policy Gradient (DDPG) technique, implemented in the Access Point (AP) to intelligently adjust the contention window. Leveraging DDPG’s capability for continuous actions, the proposed method provides Quality of Service (QoS) differentiation, ensuring that each service at its respective priority level meets its transmission requirements. Compared to the BEB algorithm, the proposed approach offers improved traffic allocation and higher network bandwidth utilization
Design of Deep Learning Acoustic Sonar Receiver with Temporal/ Spatial Underwater Channel Feature Extraction Capability
Get PDFIn this study, deep learning network technology is employed to solve the problem of rapid changes in underwater channels. The modulation techniques employed are frequency-shift keying (FSK) and the BELLHOP module of MATLAB; they are used to create water with multipath, Doppler shifts, and additive Gaussian white noise such that underwater acoustic receiving signals simulating the actual ocean environment can be obtained. The southwest coastal area of Taiwan is simulated in the manuscript. The results reveal that optimizing the environment by using the virtual time reversal mirror (VTRM) technique can generally mitigate the bit error rates (BERs) of the deep learning network’s model receiver and traditional demodulation receiver. Lastly, seven deep learning networks are deployed to demodulate the FSK signals, and these approaches are compared with traditional demodulation techniques to determine the deep learning network techniques that are most suitable for marine environments
Personalized Clothing Prediction Algorithm Based on Multi-modal Feature Fusion
Get PDFWith the popularization of information technology and the improvement of material living standards, fashion consumers are faced with the daunting challenge of making informed choices from massive amounts of data. This study aims to propose deep learning technology and sales data to analyze the personalized preference characteristics of fashion consumers and predict fashion clothing categories, thus empowering consumers to make well-informed decisions. The Visuelle’s dataset includes 5,355 apparel products and 45 MB of sales data, and it encompasses image data, text attributes, and time series data. The paper proposes a novel 1DCNN-2DCNN deep convolutional neural network model for the multi-modal fusion of clothing images and sales text data. The experimental findings exhibit the remarkable performance of the proposed model, with accuracy, recall, F1 score, macro average, and weighted average metrics achieving 99.59%, 99.60%, 98.01%, 98.04%, and 98.00%, respectively. Analysis of four hybrid models highlights the superiority of this model in addressing personalized preferences
Enhancing River Flood Prediction in Early Warning Systems Using Fuzzy Logic-Based Learning
Get PDFPrevious studies show that the fuzzy-based approach predicts incoming floods better than machine learning (ML). However, with numerous observation points, difficulties in manually determining fuzzy rules and membership values increase. This research proposes a novel fuzzy logic-based learning (FLBL) that embeds missing data imputations and a fuzzy rule optimization strategy to enhance ML performance while still benefiting from fuzzy theory. The simple moving average handles sensors’ missing data. The logical mapping is used for fuzzification automation and fuzzy rule generation. The join function between the Szymkiewicz–Simpson coefficient similarity and max function is applied to optimize a fuzzy rules model. The case study uses observation data from three rivers traversing three districts in Semarang City. As a result, FLBL achieves 97.87% accuracy in predicting flood, outperforming the decision tree (96%) and the neural network (73.07%). This work is significant as a part of preventive flood-related disaster plans
Prediction of Crop Leaf Health by MCCM and Histogram Learning Model Using Leaf Region
Get PDFThis study introduces a model called the crop leaf health prediction model (CLHPM) that utilizes a bio-inspired method to accurately identify the leaf region. This approach enhances the process of learning important features and overcomes the challenges posed by the hindrance from the chromatic and structural diversity of each leaf. To train the learning model, a modified co-occurrence matrix (MCCM) in texture analysis is used to overcome the limitations of the leaf region, and a histogram method is also deployed for color analysis. The experiment is conducted on a real dataset of tomato crop leaves. It is observed that the average accuracy has increased by 3.50%. The existing MobileNetV2 model presents an accuracy of 95.73%, and the proposed CLHPM model renders 99.23%. Moreover, an enhancement of 3.72 in the F-measure is also noticed
Current Trends in Named Entity Recognition from Automatic Speech Recognition: A Bibliometric Analysis Using Scopus Database
Get PDFNamed entity recognition (NER) is critical for language understanding and text mining systems, such as event extraction and automatic question-and-answer systems. However, NER from automatic speech recognition (ASR) outputs remains challenging due to errors and lack of textual cues. This study aims to provide a comprehensive bibliometric analysis of research on NER from ASR, focusing on publications indexed in the Scopus database before 2024 to understand the research field. Using Biblioshiny and VOSviewer tools, this research identifies the key trends, prominent authors, and international collaborations in the research network. The results show steady growth in this research area, while conference papers are the predominant source type. Additionally, the study highlights the increasing intervention of deep learning approaches to enhance NER accuracy, suggesting potential research directions to reduce error rates, and developing more robust NER algorithms. Finally, the findings underscore the importance of cross-disciplinary collaborations to document any current challenges
Domain Adaptation for Roasted Coffee Bean Quality Inspection
Get PDFCurrent research in machine learning primarily focuses on raw coffee bean quality, hampered by limited labeled datasets for roasted beans. This study proposes a domain adaptation approach to transfer knowledge acquired from raw coffee beans to the task of inspecting roasted beans. The method maps the source and target data, originating from different distributions, into a shared feature space while minimizing distribution discrepancies with domain adversarial training. Experimental results demonstrate that the proposed approach effectively uses annotated raw bean datasets to achieve a high-performance quality inspection system tailored specifically to roasted coffee beans
Analysis of Drain-Induced Barrier Lowering for Gate-All-Around FET with Ferroelectric
Get PDFThis study presents an analytical model for the drain-induced barrier lowering (DIBL) of a junctionless gate-all-around FET with ferroelectric, utilizing a 2D potential model. A multilayer structure of metal-ferroelectric-metal-insulator-semiconductor is used as the gate, as well as the remanent polarization and coercive field values corresponding to HZO are used. The DIBLs obtained with the proposed model demonstrate good agreement with those obtained using the second derivative method, which relies on the 2D relationship between drain current and gate voltage. The results demonstrate that an increase in ferroelectric thickness leads to a negative DIBL value due to the ferroelectric charge. Additionally, there exists an inverse relationship between ferroelectric thickness and channel length to achieve a DIBL value of 0. This condition is satisfied only with the increase of the ferroelectric thickness as the channel radius and insulator thickness increase. The DIBLs increase with higher remanent polarization and lower coercive field, remaining constant when the ratio of remanent polarization and coercive field is maintained
Tool Wear Prediction Based on Adaptive Feature and Temporal Attention with Long Short-Term Memory Model
Get PDFEffective monitoring of tool wear status can improve productivity and reduce losses. In previous studies, extensive feature selection was required when using the traditional machine learning method. The gating mechanism in the traditional long short-term memory (LSTM) model may incur information loss and a weaker representation of global sequential dependencies in handling long sequences. This paper aims to enhance the performance of the LSTM model in tool wear prediction by combining feature and temporal attention. Firstly, the original vibration signal is divided into sub-sequences and related features extracted. Secondly, the ability to capture global sequential dependencies using the LSTM model is improved by feature and temporal attention. Finally, a fully connected layer is used to predict tool wear values. Compared to traditional LSTM, the proposed method performs best in three evaluation metrics, RMSE, MAE, and the coefficient of determination
Enhanced Design of On-Chip Monopole Antenna Inspired by Partially Reflective Surface at 5.8 GHz
Get PDFThe increasing popularity of compact, chip-based devices has spurred interest in developing on-chip antennas (OCAs). However, OCAs suffer from low gain and poor radiation efficiency due to the silicon substrate’s low resistivity and high permittivity, influencing antenna performance. To avert these challenges, this study aims to enhance an OCA’s gain and radiation efficiency by incorporating a partially reflective surface (PRS) into the antenna structure. The antenna is simulated using 3D CST software, and its performance is evaluated. To validate the simulation, an antenna prototype is fabricated using sputtering and chemical vapor deposition (CVD) technologies. The prototype demonstrates a peak gain of 2.14 dB and radiation efficiency of 72.2%, showing a 24.3% gain increase and a 16.25% efficiency increase compared to the design without PRS. Additionally, it achieves an impedance bandwidth of 0.63 GHz, making it suitable for WiMAX, RFIC, and Wi-Fi 6 applications