Taiwan Association of Engineering and Technology Innovation: E-Journals
Not a member yet
887 research outputs found
Sort by
Enhancing Synchronization of YOLO-Based Traffic Detection on Low-End Devices by Using the COID Algorithm
Full text linkThe rapid increase in population and vehicle usage intensifies traffic congestion, creating a pressing need for accurate real-time vehicle detection. While the you only look once (YOLO) algorithm enables efficient end-to-end detection, its performance is constrained by hardware limitations, leading to desynchronization on low-end devices. To address this limitation, this study proposes the catch one image detection (COID) algorithm, which restores synchronization without altering the YOLO architecture or requiring retraining. By dynamically adjusting the frame capture interval according to hardware capability, COID ensures real-time alignment between detection and live video streams while reducing deployment complexity. Experimental evaluations on high-, mid-, and low-end devices, including validation on multi-intersection surveillance footage under low-visibility conditions, confirm the robustness and applicability of COID. These findings demonstrate COID’s practicality as a scalable solution for real-world intelligent traffic monitoring
Knowledge Representation Strategies for Reducing Hallucinations in Retrieval-Augmented Domain-Specific Question Answering
Full text linkTo address the limitations of hallucinated responses in large language models (LLMs), an artificial intelligence (AI) chatbot featuring a retrieval-augmented generation system is designed to assist with subject-based certification instruction. Focusing on the Level B certification curriculum for computer hardware repair as an example, this study develops six distinct knowledge base structures (Type0–Type5) and integrates them into two open-source 7B-parameter LLMs (LLaMA2 and Qwen2) with a custom-built question and answer system. Response accuracy to 10 standardized questions is evaluated by domain experts. Knowledge structure significantly affects performance, with the enriched Type5 base yielding the highest accuracy (Qwen2: 98 points; LLaMA2: 73 points). Statistical tests confirm significant improvements with knowledge base enhancement across knowledge types and between models. These findings highlight the critical role of knowledge representation and LLM selection in domain-specific AI applications, proffering practical design guidelines for intelligent teaching assistants in technical education
A Human Trial Study on Overcoming Linguistic and Stress-Related Barriers in Chronic Disease Management
Full text linkThe study addresses language and stress barriers in chronic disease management among Sinhala-speaking patients, using an Augmented Reality (AR) embedded platform combined with an Artificial Intelligence (AI) chatbot. Rasa is integrated into the AI chatbot using Bidirectional Encoder Representations from Transformer (BERT)-Sinhala embeddings to deliver culturally tailored, real-time health recommendations. An AR module on AR.js and Three.js implements personalized stress-reduction interventions. The chatbot achieves an F1 score of 91% and a word error rate (WER) of 8.5%, while the AR system reduces systolic blood pressure by 8.96% (p = 0.002). The combined platform attains 88% scenario accuracy with a mean response time of 1.2 seconds. These findings support the system’s potential to improve healthcare access and reduce stress markers among Sinhala-speaking chronic patients, offering a low-resource, scalable, and culturally appropriate model for healthcare environments
Wetting-Drying Durability of Lateritic Soil Stabilized with One-Part High-Calcium Fly Ash Geopolymer
Full text linkThis study investigates the durability under wetting and drying conditions of marginal lateritic soil (MLS) stabilized with a one-part high-calcium fly ash geopolymer (OPFAG). The variables include an MLS: fly ash ratio of 70:30, solid sodium hydroxide content ranging from 0 to 40%, and the number of wet-dry cycles. Durability is evaluated by measuring the unconfined compressive strength (UCS) of MLS samples stabilized with OPFAG and MLS samples stabilized with ordinary Portland cement (OPC). The results show that OPFAG improved the engineering properties of MLS. The highest UCS values are achieved at 20% solid sodium hydroxide, achieving a UCS of 1889 kPa for the geopolymer-stabilized MLS and at 5% OPC for OPC-stabilized MLS (1320 kPa). The UCS of both stabilized MLS samples increases with the number of wet-dry cycles up to 6 cycles, after which a decline is observed
Development and Improvement of a Vacuum Fryer and a Centrifugal Deoiling Machine for Deep-Fried Split-Gill Mushroom Production
Full text linkThis study aims to develop a vacuum fryer and a centrifugal deoiling machine for producing deep-fried split-gill mushrooms. The vacuum fryer prototype includes a fryer, oil heater, vacuum pump, and control system, while the deoiling machine features a steel frame, external barrel, centrifugal barrel, and transmission unit. Tests are conducted by frying split-gill mushrooms at 60, 80, and 100 ℃ for 5, 10, and 15 minutes. The deoiling machine removes oil at three rotational speeds (100, 300, and 500 rpm) and deoiling times (1, 3, and 5 minutes). Results show that the ideal frying condition is 80 ℃ for 10 minutes, and the optimal deoiling is achieved in 5 minutes at 500 rpm
Analysis of Stress and Strain in Sandwich Structures Using an Equivalent Finite Element Model
Full text linkThe study aims to build an equivalent 2D model as an alternative to the 3D model of sandwich panel structures. This model enables for reducing model building time and calculation time in the design calculation of this sandwich structure. The research object in this study is corrugated core cardboard. First, the isotropic plasticity equivalent (IPE) model for the paper material is implemented in the Abaqus software, using the VUMAT user subroutine. Subsequently, the homogenization method is proposed as an equivalent elastic-plastic finite element model. This model is implemented in Abaqus using the UGENS subroutine. Finally, numerical simulations of different load cases between the 3D model and the equivalent 2D model are performed to confirm the accuracy of the proposed model. The comparison results indicate that the equivalent model ensures exceptional accuracy compared to the 3D model but significantly reduces model building time and CPU time
Design Modification and Analysis of Brushless Direct-Current Axial Fan Motor Stator Using Taguchi and One-Factor-At-A-Time Method
Full text linkThis paper introduces a novel approach utilizing Altair Flux software for electromagnetic finite element simulation and analyses of cogging torque and back electromotive force (BEMF) without altering the rotor conditions. This investigation aims to understand the effects on the design of brushless direct current (BLDC) axial fan motors. The Altair HyperStudy optimal software is used to conduct the Taguchi experimental method to analyze the influence of critical factors in motor stator design. Subsequently, the one-factor-at-a-time method proposed in this paper is applied to find the optimal motor geometry stator design satisfying the requirements. Eventually, an experimental motor is established and compared with a commercial 9SG5748P5G01 BLDC axial fan motor. The BEMF is resultantly smaller than the 9SG5748P5G01 motor. Furthermore, the disparity between the experimental and simulation analysis results is minimal with consistent findings. The motor stator design and simulation analysis methods can potentially support the motor design
Efficient Classification of Power Quality Using Long Short-Term Memory Network Technique
Full text linkThis study aims to apply a new deep-learning technique to detect and categorize individual and complex PQ issues such as swell, flickers, surges, interruptions, and sags. The suggested technique, the long short-term memory (LSTM) network, is a novel artificial intelligence technique and an identifiable form of recurrent neural network. This technique is utilized to detect and identify power quality (PQ) issues based on three principal solutions: automatic feature extraction, voltage/current magnitude calculations, and PQ problem duration. Simulated PQ problems generated by the Matlab simulation and many real field data sets are used to authorize the proposed technique's capability. The real data contain voltage and current waveforms that are measured, recorded, and analyzed in medium-voltage and high-voltage (MV/HV) substations by using a data acquisition device. The simulation results show that the proposed method is capable of detecting and classifying PQ problems more accurately compared with other artificial intelligence techniques
Predicting Punching Shear Strength of RC Interior Flat Slabs Using an Attention-Based Transformer and Differential Evolution
Full text linkPunching shear strength (PSS) prediction in reinforced concrete (RC) interior flat slabs remains challenging for conventional empirical methods. Thus, this study proposes a hybrid attention-based transformer model optimized via differential evolution to address this limitation. The methodology combines multi-head attention mechanisms to capture nonlinear interactions among critical parameters (support dimensions, slab thickness, concrete strength, rate of steel bars, and steel yield strength). The model is trained and tested using 417 experimental results of flat slabs and processes eight input features encompassing geometric properties. The experimental results show superior accuracy, achieving a mean squared error (MSE) of 0.00017, a root mean squared error (RMSE) of 0.0113, and an R-squared of 0.998. The proposed model is benchmarked with well-known machine learning (ML) models and achieves a superior performance. These results emphasize the model’s potential as a scalable and precise tool for predicting PSS in RC flat slabs
DBSCAN-Based Minimum Enclosing Ellipse Using the Control Barrier Function for Safe Navigation of Mobile Robots
Full text linkThis paper aims to reduce the redundant unsafe area in quadratic program approaches based on the Control Barrier Function (CBF) and the Control Lyapunov Function (CLF) for collision avoidance, hereafter referred to as the CBF-CLF approach. Existing CBF-CLF quadratic program approaches typically construct CBF based on Euclidean distance; however, the redundant unsafe area due to obstacles is excessively large, which may prevent finding feasible solutions. To address this issue, this study employs density-based spatial clustering of applications with noise (DBSCAN) and the Minimum Enclosing Ellipse (MEE) to reduce the unsafe area. The proposed approach is referred to as the DBSCAN-MEE-CBF. The effectiveness of the proposed method is demonstrated through both computer simulations and real-world experiments. Specifically, the proposed method reduces the size of the redundant unsafe area by up to 26.52% while maintaining robust collision avoidance