Akademik Veri Yönetim Sistemi
Electrical Engineering, cilt.107, sa.11, ss.14827-14848, 2025 (SCI-Expanded)
Ensuring the reliable operation of electrical machines is critical for industrial efficiency, with predictive maintenance playing a key role in early fault detection. Vibration signal analysis is a widely used diagnostic tool due to its sensitivity to mechanical anomalies. In this study, we explore the impact of Generative Adversarial Networks (GANs) and Conditional GANs (cGANs) for data augmentation in fault classification using high-capacity 3D ResNet architectures. The one-dimensional signals were transformed into 3D representations by replicating the data across all channels to make vibration signals compatible with 3D ResNet input requirements. Although vibration signals are inherently 1D, the use of 3D ResNet allows for deeper spatial feature extraction across multiple axes, enabling the model to learn more complex patterns that may emerge in the augmented input space. This method takes advantage of the expressive capabilities of 3D convolutions to improve classification accuracy. Using the CWRU and HUST datasets, we augment the data by 20% and evaluate ResNet101 and ResNet152 models on both original and augmented datasets. Our results show that GAN and cGAN augmentation significantly improve classification accuracy in most cases. For instance, the F1 scores of the ResNet101 model on the CWRU dataset have increased from 88.44 to 94.76% (using GAN) and 97.13% (using cGAN), while the F1 scores on the HUST dataset have enhanced from 94.33 to 99.96% and 99.75%, respectively. To the best of our knowledge, this is the first study to integrate GAN and cGAN-based augmentation with high-capacity 3D ResNet architectures applied to pseudo-3D representations of 1D vibration signals. By transforming one-dimensional signals into three channels and facilitating deep spatiotemporal feature extraction via 3D convolutions, our methodology demonstrates notable accuracy enhancements compared to conventional 1D CNNs and earlier reported models augmented with GAN. These findings highlight the effectiveness of advanced data augmentation in improving generalization and robustness, making deep learning models more reliable for predictive maintenance applications.