Konghuayrob, Poom

Preferred name

Konghuayrob, Poom

Main Affiliation

School of Engineering

poom.ko@kmitl.ac.th

Now showing 1 - 10 of 15

Some of the metrics are blocked by your
consent settings
Comparison of Reduced-Length FFT-Based Feature for Induction Motor Fault Classification
(2025-01-01)
Taweewat, Pat
;
Suwan-Ngam, Warachart
;
Songsuwankit, Kanoknuch
;
Konghuayrob, Poom
This research presents a comparison of FFT-based features which can be used for classifying induction motor faults via neural network. In this paper, the misalignment and rotor bar damage faults are investigated by using stator current as input data only. As the length of the full FFT can include both informative data corresponding to the faults and uninformative data such as noise from environment or electrical supply, only relevant magnitude from FFT bins should be selected and used instead. This paper proposed to use threshold level determined from the magnitude of FFT bins in dataset as a criterion for the selection. From experimental results, an input feature vector created by proposed method can create short input feature vector length to be used by neural network efficiently. The trained neural network performs classification task at 99.98% in accuracy. Comparing to using dimension reduction by PCA, thresholding method needs basic computation, and yields result close to PCA method.
Some of the metrics are blocked by your
consent settings
Mitigating Racial Bias in Skin Lesion Classification With a Novel Deep Learning-Driven Dataset
(2025-01-01)
Joshi, Priyanka
;
Ruangrit, Chanet
;
Konghuayrob, Poom
(1) Background: A critical issue in the application of Machine Learning (ML) in dermatology is the presence of racial bias in training datasets, which can lead to disparities in diagnostic performance across different skin tones. This paper proposes a mitigation strategy for the BCN20000 dataset, addressing its limitations regarding darker skin lesions by augmenting it with a custom dataset. (2) Method: This was achieved by augmenting the BCN20000 dataset with a custom-developed dataset of dark skin lesion images, generated by combining serial style transfer and latent diffusion-based upscaling. The impact of this augmented dataset was assessed on the accuracy of several established Convolutional Neural Network (CNN) architectures, including DenseNet, ConvNeXt, EfficientNet, RegNet and ResNet. (3) Results: Our findings underscore the necessity of diverse datasets for AI-driven dermatology tools with the best performing model achieving 92% accuracy when trained on the augmented dataset containing additional dark-skin images compared to baseline models trained solely on BCN20000. These results align with prior studies emphasizing the critical role of representative training data in mitigating racial bias in medical AI systems.
Some of the metrics are blocked by your
consent settings
Optimizing Wafer Classification in Industrial Manufacturing Using Particle Swarm Optimization and Deep Learning
(2025-01-01)
Suwannoot, Pisit
;
Konghuayrob, Poom
In this study, we examine the application of convolutional neural networks (CNNs) for wafer pattern classification, with a focus on enhancing training efficiency and model performance. To achieve this, particle swarm optimization (PSO) is employed to improve the model performance while reducing its complexity, a critical factor in production environments. By minimizing the number of layers, the proposed method accelerates training, reduces resource consumption, and enhances defect detection accuracy. Wafer failure patterns are classified into four categories: vertical, rectangular, edge, and horizontal. The approach achieves an impressive F1-score of 0.988, significantly surpassing the traditional CNN’s score of 0.83. By integrating PSO, the method considerably improves the visual inspection process for hard disk drives, contributing to high-quality production. This optimization not only streamlines workflows but also enables manufacturers to address issues more rapidly, aligning with Industry 4.0’s objectives of automation and intelligent monitoring.
Some of the metrics are blocked by your
consent settings
Two-Tier AI Surveillance: Enhancing Weapon Detection Through Edge and Cloud Collaboration
(2026-01-01)
Sun, Sanchai
;
Songsuwankit, Kanoknuch
;
Thamrongrongveerachart, Chawapon
;
Yanyong, Sarucha
;
Sri-on, Jiramate
This research presents a novel two-tier AI-based surveillance system designed for real-time weapon detection to enhance security and prevent potential robberies. The system leverages the computational capabilities of both edge and cloud resources, integrating a YOLOv5s model on a Jetson Nano for initial detection and a YOLOv8l model on a server equipped with an Nvidia A100 for refined analysis. The initial detection performed by the Jetson Nano rapidly identifies potential threats and forwards compressed images to the server, optimizing bandwidth usage and transmission speed. Upon receipt, the server applies image enhancement techniques to restore and upscale the images from 640 pixels back to 1280 pixels before further verification. Utilizing Python Django, the server processes the enhanced images with a more sophisticated model to ensure high accuracy in detection. The integration of edge computing optimizes the system’s performance by enabling real-time processing and reducing latency. This hybrid approach enhances the efficiency and scalability of the surveillance system, ensuring robust and timely detection. Upon confirming the presence of a weapon, the system sends immediate alerts via LINE Notify to both users and local law enforcement, thereby enabling prompt responses to potential security threats. Additionally, the Jetson Nano hosts a Flask application, allowing users to download previously recorded videos for further review and evidence collection. The proposed solution combines edge computing, cloud processing, and image optimization techniques to provide an efficient, scalable, and high-performance solution for real-time weapon detection which enhancing accuracy and reliability in real-world surveillance.
Some of the metrics are blocked by your
consent settings
Decentralised and Centralised Fixed Structure H∞ Robust Loop Shaping for the MIMO Microsurgical Manipulation Based on PSOGSA
(2026-01-01)
Kaitwanidvilai, Somyot
;
Konghuayrob, Poom
;
Aoyama, Hisayuki
This paper presents decentralised and centralised fixed-structure H∞ robust control methods optimised by the Particle Swarm Optimisation and Gravitational Search Algorithm (PSOGSA) for a coupled multi-input multi-output (MIMO) microsurgical manipulator. The design framework explicitly considers uncertainties and disturbance constraints. Conventional H∞ loop-shaping controllers are typically of high order, complex, and difficult to implement in practice. To address this limitation, Proportional-Integral-Derivative (PID)-structured decentralised and centralised H∞ controllers are proposed, providing compact structures while retaining robustness. The novelty of this work lies in embedding H∞ robustness criteria into practical PID-based frameworks, bridging the gap between theoretical robust design and experimental implementation in microsurgical applications. The proposed controllers are evaluated against a reduced-order H∞ controller derived from Hankel norm approximation and a Ziegler–Nichols tuned PID controller, using both simulation and experimental studies. Results demonstrate that the proposed controllers achieve improved stability margins (0.449–0.521 compared with 0.436 for the reduced-order design), maintain low root-mean-square errors (≈0.067–0.089) and remain robust under voltage disturbances where conventional PID control fails. These findings confirm the contribution of a practical and efficient robust control strategy for enhancing the precision and reliability of microsurgical manipulators.
Some of the metrics are blocked by your
consent settings
Optimization of 3D SLAM and Localization for Quadruped Robot Dogs Using Autoencoder-based Techniques
(2025-01-01)
Thamrongveerachart, Chawapon
;
Thepsit, Thitipong
;
Ningnoi, Paphop
;
Konghuayrob, Poom
This paper presents an optimized 3D SLAM and localization framework for quadruped robot dogs using a custom autoencoder-based point cloud compression technique. Our approach combines denoising, voxelization, self-attention, and PointNet modules to reduce the size of the point cloud by more than 80% while removing outliers and maintaining structural clarity. We use a 3D laser-scan matcher for localization, which combines onboard odometry from the Unitree Go2 quadruped robot with LiDAR odometry from a Livox Mid-360 sensor. Our approach maintains accurate pose estimation and considerably reduces processing time by working directly on compressed point clouds. Comparative tests show that the system is appropriate for real-time applications in unstructured contexts due to its enhanced map clarity and decreased localization delay. The proposed framework offers an effective and computationally efficient solution for deploying 3D SLAM in agile legged robots operating in dynamic or hazardous settings.
Some of the metrics are blocked by your
consent settings
Hybrid Deep Learning and FAST-BRISK 3D Object Detection Technique for Bin-picking Application
(2024-01-01)
Taweesoontorn, Thanakrit
;
Yanyong, Sarucha
;
Konghuayrob, Poom
In the field of industrial robotics, robotic arms have been significantly integrated, driven by their precise functionality and operational efficiency. We here propose a hybrid method for bin-picking tasks using a collaborative robot, or cobot combining the You Only Look Once version 5 (YOLOv5) convolutional neural network (CNN) model for object detection and pose estimation with traditional feature detection based on the features from accelerated segment test (FAST) technique, feature description using binary robust invariant scalable keypoints (BRISK) algorithms, and matching algorithms. By integrating these algorithms and utilizing a low-cost depth sensor camera for capturing depth and RGB images, the system enhances real-time object detection and pose estimation speed, facilitating accurate object manipulation by the robotic arm. Furthermore, the proposed method is implemented within the robot operating system (ROS) framework to provide a seamless platform for robotic control and integration. We compared our results with those of other methodologies, highlighting the superior object detection accuracy and processing speed of our hybrid approach. This integration of robotic arm, camera, and AI technology contributes to the development of industrial robotics, opening up new possibilities for automating challenging tasks and improving overall operational efficiency.
Some of the metrics are blocked by your
consent settings
Prediction of Flying Height Using Deep Neural Network Based on Particle Swarm Optimization in Hard Disk Drive Manufacturing Process
(2024-01-01)
Kanjanapruthipong, Worawit
;
Prasitmeeboon, Pitcha
;
Konghuayrob, Poom
In contemporary hard disk drive (HDD) manufacturing processes, after the assembly of the HDD from the production line, a series of diverse calibration procedures are necessary to ensure standardization. These include capacity calibration, which determines the storage space in terabytes (TB) presently available, and flying height (FH) calibration, which evaluates the distance between the head and the disk by applying electric current to the heater coil element to achieve the desired FH, thus optimizing the writing and reading performance and tailoring it to each HDD. Additionally, electric current is saved in a digital-to-analog converter (DAC) unit for the utilization of a read/write head, while a preamp collaborates with the drive firmware to convert the electric current in the DAC unit to milliwatts. In the present scenario, multiple calibrations of flying heights (FHs), specifically flying height 1 (FH1) and flying height 2 (FH2), are performed. Each FH calibration requires a testing time of approximately 5 h owing to the separation of measurement points into 240 locations across the disk surface, referred to as test zones, with a total of 20 heads. The primary objective of this study is to reduce the testing time by using a combination of deep neural network (DNN) and particle swarm optimization techniques to predict the DAC profiles of FH2 as it approaches FH1, where FH1 is the input for the DNN model.
Some of the metrics are blocked by your
consent settings
A Small Deep Learning Model for Fault Detection of a Broken Rotor Bar of an Induction Motor
(2024-01-01)
Taweewat, Pat
;
Suwan-Ngam, Warachart
;
Songsuwankit, Kanoknuch
;
Konghuayrob, Poom
In this paper, we present an investigation of a small deep learning model applied to the detection of a broken rotor bar of an induction motor. The motor current spectrum analysis is the base method for fault detection. This proposed method focuses on the analysis of the modification of the input vector and model configuration. This method was implemented and it showed that the feature length and size of the model are reduced compared with the existing method. The experimental results showed that only feature extraction using the spectral-based method and limit range of its coefficient are adequate to provide accuracy of small deep learning comparable to that of the parallel-layer deep learning model. Likewise, at the same accuracy level, based on the deep learning model, a shorter sampling duration than that required by the reference model is needed.
Some of the metrics are blocked by your
consent settings
Anomaly Flying Height Prediction Based on Clustering Techniques in Hard Disk Drive Manufacturing
(2025-01-01)
Kanjanapruthipong, Worawit
;
Konghuayrob, Poom
In this research, we present a method for predicting anomaly flying height (FH) profiles in hard disk drive (HDD) manufacturing by analyzing FH data at the FH1 stage. Anomalies at FH1 can lead to calibration issues at FH2, disrupting the production process. We propose an AI-based approach using unsupervised clustering techniques to group FH profiles of the read/write head. We evaluated four clustering algorithms, KMeans, MiniBatchKMeans, Birch, and BisectingKMeans, along with the Elbow method to determine the optimal number of clusters. By identifying anomalous FH profiles early at FH1, the method enables proactive intervention, reducing calibration process time and improving production efficiency. Our model achieved an accuracy of 0.939 without relying on manual feature selection (e.g., pressure and temperature), which is often difficult to capture using traditional linear or rule-based models owing to the nonlinear nature of FH profiles. These results demonstrate the practical potential of clustering techniques in enhancing HDD manufacturing processes.

Konghuayrob, Poom

Filters

Settings

Sort By

Results per page