Akademik Veri Yönetim Sistemi
5TH INTERNATIONAL CONFERENCE ON INNOVATIVE ACADEMIC STUDIES ICIAS 2024, Konya, Türkiye, 10 - 11 Ekim 2024, ss.245-250
This study focuses on fault detection in
electrical machines using a one-dimensional Convolutional Neural Network
(1D-CNN) model. Two publicly available datasets, CWRU and HUST, are utilized to
evaluate the model’s performance with datasets of varying sizes and
complexities. The CWRU dataset includes four fault categories: Normal, Ball,
Inner Race, and Outer Race, while the HUST dataset features seven fault
categories: Normal, Ball, Inner Race, Outer Race, combinations of Inner Race
and Ball, Outer Race and Ball, and Inner Race and Outer Race faults. Both datasets
are divided into training, validation, and test sets, following a standardized
split—20% for testing, and 20% of the remaining data for validation. The model
is trained and validated for 30 epochs using the separated datasets. The
results show that when using the CWRU dataset, the gap between training and
validation accuracy is larger, suggesting that the model is overfitting and not
generalizing well to unseen data. In contrast, the HUST dataset results in a
smaller accuracy gap, indicating better generalization and performance
consistency across training and validation. This suggests that the CWRU dataset
may contain more complexity, noise, or class imbalance compared to the HUST
dataset, making it challenging for the model to generalize effectively.
Consequently, the model achieves an overall accuracy of 85.68% on the CWRU
dataset and 98.41% on the HUST dataset during testing, aligning with the
predictions made from the accuracy-loss plots. This study concludes that
dataset selection is crucial for achieving optimal performance in deep learning
models for fault detection. The results highlight the importance of evaluating
model performance on different datasets to ensure robust and generalized fault
detection capabilities.