Browsing by Author "Yetkin, E.F."

Now showing 1 - 3 of 3

Citation Count: 0
Automatic Segmentation of Time Series Data With Pelt Algorithm for Predictive Maintenance in the Flat Steel Industry
(Institute of Electrical and Electronics Engineers Inc., 2024) Kaçar, S.; Balli, T.; Yetkin, E.F.
In this study, we aim to test the usability of Change Point Detection (CPD) algorithms (specifically the Pruned Exact Linear Time-PELT) to facilitate the utilization of large volumes of data within predictive mechanisms in the industry. We proposed an efficient CPD parameter selection mechanism for defect diagnosis using time-series vibration data from critical assets. We emphasized the practical algorithm PELT to ensure broad industrial applicability. Our experimental analysis, using synthetic and actual vibration data, demonstrated the practical applicability and effectiveness of PELT algorithm for automatic segmentation. The numerical results show the potential of CPD methodologies for improving predictive maintenance operations by providing an automatic segmentation mechanism. This pipeline proposes a way to increase the operational efficiency and scalability of predictive maintenance approaches, enhancing maintenance procedures and ensuring the long-term reliability of industrial systems. © 2024 IEEE.
Citation Count: 0
Comparison of Feature Selection Methods for Mechanical Properties of Cold Rolled Products in Flat Steel Manufacturing
(Institute of Electrical and Electronics Engineers Inc., 2024) Ilme, D.B.; Öper, M.; Yetkin, E.F.
The mechanical properties of steel are critical for ensuring its quality and are traditionally tested using destructive methods, which involve cutting test samples after the skin-rolling process. This procedure necessitates the scrapping of the last 8 meters of the coil and extracting a 500 mm wide sample, consuming approximately 1 to 1.5 minutes. To eliminate these additional process steps and minimize material waste, this study aims to predict steel coils' yield strength and tensile strength in the flat steel industry using six machine learning models. The models incorporate 24 distinct production parameters as inputs. The models examined include Linear Regression, Support Vector Regressor (SVR), Decision Tree, K-Nearest Neighbors (KNN), Random Forest, and eXtreme Gradient Boosting (XGBoost). To enhance the predictive performance of these models, seven different feature selection methods are employed. These methods systematically rank the production parameters based on their influence and are iteratively utilized within the models to refine their accuracy. The application of these feature selection techniques significantly improves the models' efficiency, leading to substantial operational benefits. The study demonstrates that machine learning models, when optimized with advanced feature selection methods, can accurately predict the mechanical properties of steel, thereby reducing the need for destructive testing. This approach not only conserves material and time but also enhances the overall efficiency of the production process in the flat steel industry. © 2024 IEEE.
Citation Count: 0
On Symbolic Prediction of Time Series for Predictive Maintenance Based on Sax-Lstm
(Institute of Electrical and Electronics Engineers Inc., 2024) Güler, A.; Balli, T.; Yetkin, E.F.
This work proposed a new forecasting approach for predictive maintenance in industrial settings, combining standard segmentation approaches like Symbolic Aggregate Approximation (SAX) and Piecewise Aggregate Approximation (PAA) with LSTM (Long-Short Time Memory). The work aims to construct a robust forecasting mechanism to estimate maintenance requirements in advance properly. We first demonstrated the results of the proposed approach for synthetically generated data and extended the results with real industrial vibration data. The algorithm's performance is assessed using real-world industry data from steel production furnaces, where timely maintenance is critical for increasing operating efficiency and reducing downtime. Experimental results show that using SAX and LSTM for forecasting industrial time series data achieves high accuracy rates (90.2 %) in a reasonable computational time. © 2024 IEEE.