CapsPDNet: Optimized Capsule Network for Predicting Insulator Discharges Using UHF Signals
IEEE Transactions on Instrumentation and Measurement, cilt.74, 2025 (SCI-Expanded, Scopus)
- Yayın Türü: Makale / Tam Makale
- Cilt numarası: 74
- Basım Tarihi: 2025
- Doi Numarası: 10.1109/tim.2025.3562974
- Dergi Adı: IEEE Transactions on Instrumentation and Measurement
- Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Academic Search Premier, Aerospace Database, Applied Science & Technology Source, Business Source Elite, Business Source Premier, Communication Abstracts, Compendex, Computer & Applied Sciences, INSPEC, Metadex, Civil Engineering Abstracts
- Anahtar Kelimeler: Capsule network, deep learning (DL), discone antenna, partial discharge (PD), predictive maintenance
- Süleyman Demirel Üniversitesi Adresli: Evet
Özet
Outdoor insulation systems often suffer from partial discharge (PD) faults that compromise the reliability of electrical grids. This paper proposes an integrated approach combining a non-invasive downsized discone antenna to capture ultra-high-frequency signals and an optimized capsule network (CapsPDNet) for robust PD fault prediction. The CapsPDNet is chosen as it overcomes the information loss associated with pooling operations and leverages vector representations, providing more nuanced predictions than scalar values, especially when the predictions are made across different antenna locations. Three typical PD fault types are targeted, surface discharges on ceramic and polymeric materials, internal discharges, and corona discharges, and then multiple signal processing techniques are evaluated to determine the most effective feature extraction and reconstruction technique. Experimental results show that the proposed model achieves up to a 18.89% improvement in PD fault prediction accuracy over benchmark approaches, demonstrating high generalizability and scalability when considering different antenna locations. The patented antenna used in this study offers a compact size through a novel size reduction technique, covers a wide bandwidth, and exhibits high sensitivity to PD activities. Additionally, discrete wavelet decomposition (DWT) is chosen as the feature extractor as it provides high-importance features with the lowest computational time, offering a more reliable and scalable solution for early PD fault detection and real-time condition monitoring of insulation systems.