| Title |
Anomaly Detection of OLTC using SE-CSFlow Network with Acoustic Data |
| Authors |
김동현(Dong Hyun Kim) ; 황호성(Ho Seong Hwang) ; 김호철(Ho Chul Kim) |
| DOI |
https://doi.org/10.5573/ieie.2025.62.5.46 |
| Keywords |
Anomaly detection; Factory facility; One-class learning; OLTC; Acoustic data |
| Abstract |
In this paper, we proposes a robust anomaly detection method based on one-class learning using acoustic data for detecting faults in On-Load Tap Changers (OLTC). The acoustic data is preprocessed into a 2D scalogram using Discrete Wavelet Transform (DWT) to enable time-frequency analysis. The proposed network is built upon CSFlow, which integrates Normalizing Flow with a memory-efficient embedding approach, and incorporates a Squeeze-and-Excitation (SE) Block to enhance feature extraction and anomaly detection performance. Experimental results on an OLTC acoustic anomaly detection dataset demonstrate that the proposed SE-CSFlow network achieves the best performance by maintaining zero Escape cases while minimizing Over Kill to 2,333 cases. These results validate the effectiveness of the proposed method in diagnosing OLTC faults even in environments with limited fault data, contributing to improved system reliability and operational stability. |