Engineering Journal (Faculty of Engineering, Chulalongkorn University, Bangkok)
Not a member yet
1223 research outputs found
Sort by
An Innovative Machine Learning Framework for Precise Quantification of Leaf Disease Severity via Proposed Hybrid Segmentation and Color Manipulation Technique
No full textBackground: Precise quantification of leaf disease severity is crucial for effective plant pathology and accurate disease management. Existing segmentation techniques, including deep learning-based methods, face challenges with noise, artifacts, and occlusions in leaf images, leading to unreliable segmentation and severity assessments. Additionally, visually similar symptoms across different diseases make accurate differentiation challenging, especially when models lack training on diverse datasets. Complex backgrounds in infected leaf images further exacerbate these limitations.
Objective: This study proposes a novel segmentation and quantification approach to overcome the limitations of existing methods, accurately extracting lesion areas and quantifying both lesion and healthy leaf pixels to measure infection severity.
Methods: The proposed approach integrates four distinct algorithms. Algorithm 1 utilizes Adaptive K-Means Clustering to create a background mask, effectively isolating leaf pixels. Algorithm 2 applies Gaussian Mixture Models (GMM) in the Lab* color space to refine the foreground mask, accurately identifying lesion regions. Algorithm 3 employs advanced color manipulation for precise lesion delineation, while Algorithm 4 combines the outputs to determine the total pixel count of leaf and lesion areas, enabling a reliable estimation of disease severity.
Results: The method was validated on a dataset of 50 images, including tomato, paddy, cucumber, and wheat leaves affected by Late Blight, Brown Spot, Downy Mildew, Septoria, and Stripe Rust. Metrics like Dice Coefficient, F1-Score, Jaccard Index, and Latency assessed segmentation quality, while Mean Absolute Error (MAE) and correlation evaluated disease severity quantification. The proposed approach outperformed state-of-the-art techniques, demonstrating superior accuracy.
Conclusion: The improved precision achieved with the proposed method supports more reliable disease assessment, aiding in the development of effective crop protection strategies and enhancing decision-making in plant disease management
Accounting of the Overburden Pressure on Analysis of Down-Hole Seismic Shear Wave Velocity
No full textShear wave velocities in soil reveal vital information to geotechnical engineers and their rates directly relate to effective stress conditions. Studies have extensively validated this dependency in theoretical models and laboratory tests but field measurements across different geographical regions require further exploration. This research analyzed seismic down-hole measurements across three Thai locations including Bangkok and the Maptaphut Industrial Estate and Chiang Mai province where the soils and geological landscapes differed substantially. The study evaluated four computational methods to calculate shear wave velocities including time-depth plot analysis and depth-time difference ratio and trigonometric distance-time difference approach and the virtual interface method that handles refraction effects at measurement depths. Traditional methods showed reduced capability to detect changes in overburden pressure because they produced sharp velocity fluctuations in shallow zones and at layer boundaries. The virtual interface method delivered superior velocity profile accuracy by integrating overburden pressure effects in the analysis. The site-specific logarithmic patterns between shear wave velocity and depth variations produced coefficients K and m that reflect soil composition and effective stress levels respectively. The studied regions exhibited contrasting parameters as Maptaphut sandy soils presented higher K values (100-160) and depth-dependent m values while Bangkok clay demonstrated K=12.52 and m=0.90. The trigonometric distance-time difference method stood out as an efficient computational method which minimized errors yet kept its applications straightforward
Comparative Analysis of CVaR Newsvendor and Traditional Inventory Management Models: An Empirical Study in the Electric Vehicle Industry
Full text linkThis comparative study examines inventory decision-making performance using the Conditional Value at Risk (CVaR) extended newsvendor model versus the traditional newsvendor framework. The analysis employs actual monthly sales data from the electric vehicle industry across diverse market scenarios. Electric vehicle supply chains face significant challenges from demand volatility and tail risk—the potential for extreme events that adversely affect supply chain operations. This research evaluates how the two models perform in the face of these market challenges, providing insights to inform improved inventory decisions in unpredictable environments. The paper introduces the CVaR criterion, along with risk and loss-aversion parameters derived from traditional models, to enhance risk-control capabilities and decision robustness. Three distinct market scenarios—normal, high-demand, and low-demand conditions—were established. Python was used to determine optimal order quantities, complemented by a sensitivity analysis examining parameter variations. Findings indicate that the CVaR-extended newsvendor model outperforms traditional approaches, particularly in terms of fitting accuracy. The model demonstrates effectiveness in reducing fluctuations in ordering strategy during peak-demand scenarios, as measured by the root-mean-square absolute error (RMSAE), and in maintaining stability under extreme conditions. This capability substantially enhances the management of both stockout and inventory backlog risks. The model demonstrates validated applicability and robustness within high-uncertainty environments characteristic of risk-intensive sectors. Research findings establish a more adaptable optimization framework for inventory management, with practical significance for industries facing challenging risk conditions
Investigating the Predictability of Contralateral Papillary Thyroid Carcinoma: An Empirical Study Using Machine Learning and Statistical Models
Full text linkPapillary thyroid carcinoma (PTC) is the most common subtype of thyroid malignancy and is increasingly diagnosed worldwide. Although typically localized, PTC exhibits a notable risk of bilateral involvement, with contralateral disease occurring in up to 44 percent of cases. While completion thyroidectomy is recommended in selected high-risk scenarios, it carries potential complications, making accurate prediction of contralateral involvement essential. This study investigates the utility of classical statistical methods and machine learning (ML) algorithms—Support Vector Machines (SVM), Extreme Gradient Boosting (XGBoost), and Random Forests (RF)—in predicting contralateral PTC using a retrospective dataset of 122 lobectomy patients. Clinical and pathological features such as tumor size, multifocality, extrathyroidal extension, and histological subtype were analyzed, with class imbalance addressed using SMOTE, ADASYN, and Borderline-SMOTE techniques. Logistic regression identified multifocality as a risk factor and revealed that certain histological variants, particularly infiltrative and invasive encapsulated follicular variants, may be protective. ML models, especially RF with Borderline-SMOTE, showed the best predictive performance. Although model accuracy remains moderate, the findings support the integration of statistical and ML approaches to enhance risk stratification and guide personalized surgical planning in PTC management
A Novel Approach for Fabricating High-Performance Aluminum Matrix Composites: Friction Stir Processing with Micro-TiO2 Particle Reinforcement by Grey-Taguchi Multi-Response Optimization
Full text linkThe development of high-performance Aluminum Matrix Composites (AMCs) is critical due to the increasing demand for materials with superior mechanical properties, particularly in aerospace and automotive industries. This research addresses the need to enhance the hardness and impact energy of AA6061-T6 aluminum alloy by reinforcing it with micro-TiO2 particles through Friction Stir Processing (FSP). The primary objective is to optimize the FSP parameters to improve these mechanical properties. A Grey-Taguchi method was employed for multi-response optimization, focusing on tool rotational speed, traverse speed, and TiO2 particle volume. The methodology utilized Taguchi Orthogonal Arrays (OA) to minimize experimental runs while still capturing comprehensive data. Grey Relational Analysis (GRA) was applied to handle multiple correlated responses, converting them into a unified metric, the Grey Relational Grade (GRG). The results identified the optimal FSP parameters as a tool speed of 1100 rpm, traverse speed of 20 mm/min, and TiO2 particle volume of 450 mm³, which significantly enhanced the mechanical properties. Comparative analysis revealed that the optimal parameters improved both hardness and impact energy by 15.80 J, with a GRG value of 0.905, indicating strong correlation between predicted and experimental outcomes. Confirmation experiments validated these results, with a 0.099 increase in GRG, suggesting that the combination of process parameters was highly effective. The findings highlight the significant influence of TiO2 particle volume on the mechanical performance of the composite. These results provide critical insights to produce advanced AMCs, offering pathways to achieving high-performance materials for industrial applications
Predicting Lung Cancer Using ResNet-50 Deep Residual Learning Model using Relu-Memristor Activation Function
Full text linkLung cancer continues to stand as a prominent contributor to deaths caused due to cancer throughout the world, underscoring the paramount importance of precise and timely detection methodologies. In this paper, we introduce a pioneering strategy aimed at forecasting lung cancer by utilizing the power of deep residual learning, coupled with a meticulous examination of medical images at the pixel level. The given methodology uses convolutional neural networks (CNNs), specifically the Residual Network (ResNet) architecture, to extract intricate features from lung images, while also emphasizing the importance of raw pixel values as input features. Our experimental results showcase the promise of this novel approach, demonstrating remarkable predictive performance in lung cancer detection. By incorporating pixel values and deep residual learning, our model achieves a high degree of accuracy, sensitivity, and specificity, surpassing existing methodologies. This research makes important contribution in the advancement of early detection of lung cancer, ultimately enhancing the recovery chances of patient and potentially minimizing the burden of this devastating disease
Virtual Reality Application of Lathe Machine Training
Full text linkThis study aims to utilize Virtual Reality (VR) technology to develop an industrial training application with content focusing on Machining operation using lathe machine. The focus of the study is to present the content being taught in manufacturing laboratory classes about lathe machining operations in the virtual world. Machining operations demonstrated within this application includes turning, facing, chamfering, cutoff, internal and external threading. The user journey is designed to guide users through the content tailored aimed to familiarize them with the order of operation and safety guidelines. This educational tool further provides the freedom of the user to repeat the process without restriction to their location, time, or resource, as opposed to conventional hands-on laboratory. The study involves the research of the laboratory content, development planning, assets creation, functional design and implementation, and integration with Salesforce platform as learning assessment database tool. This study emphasizes future content expansion and is therefore developed with modularity in mind to best promote future content enrichment. The application resulted in the ability to represent a total of 5 machining operations
Routing-based Optimal Topology Control for Improving Quality of Transmission and Energy Efficiency in Wireless Sensor Networks
No full textEfficient energy utilization in wireless sensor networks (WSN) is one of the factors that improves network performance. Recent works have focused on energy-aware routing protocols or topology control to improve energy efficiency at each sensor node. However, these studies separate routing and topology control. This can result in the fact that if an optimal topology is obtained, the data transmission routes may not be optimal and vice versa. To overcome this problem, we propose a routing-based topology control algorithm to obtain an optimal network topology and data transmission route set. Our idea is to use the integer linear programming (ILP) problem to formulate the routing problem with the objective function of optimizing two metrics: the distance of the routes and node degree of the topology. The constraints include the traffic load on the wireless links and desired node degree. By solving this ILP problem to determine a set of data transmission routes, the transmission power of the sensor nodes is simultaneously determined based on its routing table to optimize the network topology. The simulation results show that the proposed algorithm outperforms well-known topology control algorithms in terms of the desired node order and energy efficiency
Design and Performance Analysis of New Seventeen Level Reduced Switch Count Multilevel Inverter
No full textMultilevel inverter has emerged newly as a very important alternate in the area of high-power medium-voltage energy control. The main problem in this technology is high devices, total standing voltage, cost, THD and efficiency. This paper's main goal is to provide a 17-level multilevel inverter with only 8 switches and 4 diodes. Four different unequal DC sources are used in this arrangement to produce the output voltage waveform with 17 levels. For the proposed inverter, the cost function per level, efficiency, total standing voltage, conduction losses, and switching losses have all been calculated in detail. Typical inverters have higher switching losses, prices, and harmonic distortion because they have more semiconductor power switches, diodes, capacitors, driver circuits, and DC sources. The switching components of the suggested arrangement have been controlled using the nearest-level control approach. The suggested multilevel inverter offers a higher efficiency of 98.24%, cost function of 5.7 for a weight coefficient of 1, improved power quality, and higher reliability. The total harmonic distortion produced by this inverter is 3.43%, which comes under the IEEE standard of 5%
Permeability of Saturated Sands and Gravels: Pore Constriction Size Perspective
No full textThis study re-analyzes saturated permeability data for sands and gravels to identify the key predictors. The dataset (76 tests) is evaluated against void ratio, specific surface area, mean pore size, and mean constriction size. The strongest correlation is found with mean constriction size. Based on this finding, the validated and refined predictive permeability model is proposed using mean constriction size as a representative parameter. The uncertainty and statistical analysis of the model indicates promising accuracy and reliability in predicting permeability for the dataset. The model is applicable for soils with coefficient of uniformity (CU) up to 76 and effective particle size (D10) up to 6 mm