About this Abstract |
| Meeting |
2026 TMS Annual Meeting & Exhibition
|
| Symposium
|
Late News Poster Session
|
| Presentation Title |
H-39: Unsupervised Health Monitoring of High-Pressure Hydrogen Compressor Seals via Optimized Multi-Axis Vibration Analysis and Deep Autoencoders |
| Author(s) |
Kyung Hwan Hwang, Churl Min Kim, Byeong Chan Choi, Hyoung Chan Kim |
| On-Site Speaker (Planned) |
Kyung Hwan Hwang |
| Abstract Scope |
High-pressure (990 Bar) hydrogen compressor seal integrity is critical for safety, yet acquiring failure data is often impractical. This study presents an unsupervised anomaly detection framework using only normal operation vibration data from 3-axis accelerometers. We systematically compare time-frequency feature extraction methods coupled with corresponding Convolutional Neural Network autoencoders (1D vs. 2D). The effectiveness of single-axis, dual-axis, and tri-axial data fusion strategies, optimized via Optuna, is evaluated based on reconstruction error. While the 3-axis STFT model demonstrated the lowest error in defining the normal state, the YZ-axis STFT combination uniquely identified clustered anomalies, indicating sensitivity to specific incipient fault patterns. An ensemble approach combining FFT and STFT outputs further enhanced detection sensitivity. This comparative analysis provides crucial insights into selecting optimal signal processing and data fusion strategies for robust seal health monitoring under data scarcity conditions. |
| Proceedings Inclusion? |
Undecided |
| Keywords |
Machine Learning, Other, Other |