KMITL

Permanent URI for this communityhttps://dspace.kmitl.ac.th/handle/kmitl/1

Browse

Search Results

Now showing 1 - 6 of 6

Some of the metrics are blocked by your
consent settings
Enhancing Industrial Machine Sound Anomaly Detection Using STFT Integrated with DWT and Autoencoder-Based Neural Networks
(2026-01-23)
Mattanaweerapong, Tawan
;
Jearanaitanakij, Kietikul
This study proposes a hybrid feature extraction approach that integrates the Discrete Wavelet Transform (DWT) with the Short-Time Fourier Transform (STFT) to improve the accuracy of anomalous sound detection in industrial machines. Conventional STFT-based methods, while effective in representing time–frequency characteristics, exhibit limitations in handling non-stationary noise and transient variations, which often lead to reduced anomaly detection performance in practical industrial environments. To address this problem, the proposed method incorporates multiresolution analysis through DWT, enhancing the system’s capability to capture both spectral and temporal information with improved noise robustness. The MIMII dataset (valve,-6 dB, ID02) was used to evaluate the model, where the DWT–STFT feature, representation was applied to an autoencoder for unsupervised anomaly detection. Experimental results demonstrate that the integration of DWT effectively enhanced noise robustness and improved classification metrics, achieving higher AUC and F1-scores compared to the baseline STFT-based approach. In conclusion, the proposed DWT–STFT fusion provides a more resilient and discriminative feature representation, making it a promising technique for practical industrial anomaly detection systems.
Some of the metrics are blocked by your
consent settings
Performance Evaluation of Imputation Techniques for Telecommunications Customer Clustering
(2026-01-01)
Sukthong, Patthama
;
Charoenporn, Pattama
Missing data significantly degrades machine learning model performance in telecommunications customer analytics, leading to unreliable customer segmentation and suboptimal business decision-making. This research systematically compares seven imputation techniques across three missing mechanisms (MCAR, MAR, MNAR) and four missing rates (5%, 10%, 20%, 30%) using the Telco Customer Churn Dataset (7,043 records). Methods evaluated include traditional approaches (mean/mode, forward fill, regression), machine learning techniques (KNN, Random Forest, MICE), and deep learning (Autoencoder). We assessed model performance using normalized MAE and RMSE, and evaluated downstream effects through clustering algorithms. Results demonstrate Random Forest imputation's superior performance with MAE of 0.1568 and RMSE of 0.2123, achieving 53.7% lower error rates compared to mean/mode imputation. Statistical analysis confirmed significant performance differences (Friedman test: χ<sup>2</sup> = 55.85, p < 0.001). Interestingly, clustering performance did not directly correlate with imputation accuracy; the Autoencoder achieved the highest silhouette score (0.1510) despite moderate reconstruction accuracy. Machine learning approaches maintained robust performance across all missing data mechanisms, whereas traditional methods degraded under MNAR conditions. These findings provide evidence-based guidelines for selecting appropriate imputation techniques in telecommunications analytics, enabling improved customer segmentation and business outcomes.
1
Some of the metrics are blocked by your
consent settings
A novel approach to enhanced fall detection using STFT and magnitude features with CNN autoencoder
(2025-02-01)
Soontornnapar, Tomorn
;
Ploysuwan, Tuchsanai
The ability to accurately detect and classify falls is critical for ensuring timely medical intervention, especially for the elderly, who face a significantly higher risk of severe injuries, loss of independence, or fatal outcomes from falls. This paper introduces a novel fall detection approach that addresses these urgent needs by using short-time Fourier transform (STFT) images and the magnitude of quaternion (MQ) signals, fused into STFT-MQ images. The proposed method leverages STFT’s time–frequency representation to capture rapid changes and dynamic characteristics in human motion data from wearable sensors, enhancing its ability to distinguish between fall and non-fall incidents. Utilizing a convolutional neural network autoencoder (CNN-AE), an unsupervised learning model, this approach analyzes transformed data without extensive labeled datasets, offering a scalable solution in diverse settings. Tested on the HIFD dataset with heart rate and IMU sensor data, the STFT-MQ-AE method achieves remarkable sensitivity of 98.08%, specificity of 98.78%, and an overall accuracy of 98.57%, setting a new benchmark in fall detection accuracy. Furthermore, the model’s reliance on an N-way K-shot learning approach enables it to manage unforeseen fall cases effectively without retraining, enhancing adaptability and real-world utility. The model achieves the highest Youden’s index (YI) of 96.85%, underlining balanced performance between fall and non-fall classification. Consistent performance across varied training scenarios yields an average accuracy of 96.10%, making this approach highly reliable. This advancement in fall detection technology offers a practical, effective solution to reduce fall-related injuries and enable timely assistance, thereby promoting safer, more independent living for at-risk populations.
1
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
FAST DEFECT DETECTION FOR GLASS BOTTLE USING AUTOENCODER AND ERROR THRESHOLD
(2024-06-01)
Hanskunatai, Anantaporn
;
Jaiyen, Saichon
;
Claypo, Niphat
Glass bottle defect detection is an important part of quality control process in any glass manufacturing industry. The bottles must be inspected before packaging. Machine vision for glass bottle defect detection is the technology and method to inspect and analyze the defects for images automatically. Machine vision requires high-ability method to detect the defect and reject the bottle with the defect quickly. In this paper, defect detection framework for glass bottle defect detection tasks using autoencoders and error threshold is proposed. The fast detection method, a small autoencoder neural network architecture was designed with only good bottle images to train an autoencoder neural network. The decoded images are representations of normal bottle images and calculate threshold errors value. Defect detection is done by comparing the error between the normal background image and the encoded images to a threshold error from the training set. The performance of our method was compared to several other methods: VGG16, MobileNetV3, ADA, edge detection and image threshold. The experimental results show that our method yields 80% of accuracy on the body dataset and 92% of accuracy on the neck dataset. The average training time of our method is faster than that of all other neural network-based methods. From the experimental results, we can conclude that our defect detection framework outperforms other approaches both in accuracy and training time for defect detection on the side wall of a glass bottle.
1
Some of the metrics are blocked by your
consent settings
An extreme learning machine based pretraining method for multi-layer neural networks
(2018-07-02)
Noinongyao, Pavit
;
Watchareeruetai, Ukrit
One approach in training a deep neural network to perform effectively is to do unsupervised pretraining on each layer, followed by fine-tuning the whole network. A common way is to train an unsupervised model of neural network such as restricted Boltzmann machines or autoencoders and stack them on top of another. Although these unsupervised pretraining approaches yield good performance, relying on back-propagation, due to iterative learning process, they still suffer from a long pretraining time. Extreme learning machine (ELM) is an analytical training approach which is extremely fast and gives a solution with a good generalization performance. In this paper, we apply a new ELM based unsupervised learning, named backward ELM based autoencoder (BELM-AE), to pretrain each layer of a neural network before using a back-propagation based learning algorithm to fine-tune the whole network. Experimental results show that the new pretraining method requires significantly shorter training time and also yields better testing performance on various datasets.

KMITL

Browse

Filters

Settings

Sort By

Results per page

Search Results