Synthetic Data Generation and Validation Framework for Synthetic-to-Real Transfer Sentetikten Gerçe?ge Aktarim için Sentetik Veri Üretimi ve Dog?rulama Çerçevesi
34th Signal Processing and Communications Applications Conference, SIU 2026, İstanbul, Türkiye, 7 - 10 Temmuz 2026, (Tam Metin Bildiri)
- Yayın Türü: Bildiri / Tam Metin Bildiri
- Doi Numarası: 10.1109/siu71813.2026.11636720
- Basıldığı Şehir: İstanbul
- Basıldığı Ülke: Türkiye
- Anahtar Kelimeler: Convolutional Autoencoder, Induction Motor Fault Diagnosis, Synthetic Data Generation, Synthetic-to-Real Transfer, Time-Series Signals, WGAN-GP
- Marmara Üniversitesi Adresli: Evet
Özet
This paper proposes a synthetic data generation framework to address the limitations of small datasets in rotating machinery fault. The framework combines a convolutional autoencoder (CAE) with a Wasserstein Generative Adversarial Network with Gradient Penalty (WGAN-GP) to generate vibration and current signals. The CAE compresses raw time-series signals into a compact latent representation, while the WGAN-GP learns the latent distribution to generate realistic signal representations, which later decoded to produce synthetic time-domain signals. Statistical, time-domain, and frequency-domain analyses demonstrate a strong similarity between real and synthetic signals. Furthermore, synthetic-to-real experiments achieve performance comparable to models trained on real signals. These results indicate that the proposed framework can effectively augment limited datasets and support robust fault detection.