We found a match
Your institution may have access to this item. Find your institution then sign in to continue.
- Title
LSTM-AE based condition monitoring for reciprocating air compressors considering on/off characteristics.
- Authors
Kim, Myeong-Joon; Cho, Hyun-Jik; Kang, Chul-Goo
- Abstract
Long short-term memory-autoencoder (LSTM-AE) models are commonly used to detect anomalies in time-series data. However, conventional LSTM-AE models may show low performance in anomaly detection when the system cycles on and off aperiodically because the model cannot learn both characteristics in the same way. To improve this, we propose an on/off LSTM-AE model for anomaly detection in a reciprocating air compressor considering its on/off characteristics. Data collected from onboard air compressors in a Seoul subway train in Korea were divided into two sets: 'on' and 'off' data based on valve signals. Separate LSTM-AE models were trained for each set. The conventional and proposed on/off LSTM-AE models were trained using normal data, and their performances were verified with artificially generated abnormal data. Anomaly scores were calculated for anomaly detection and the performance of the models was evaluated utilizing the F1 score. The proposed on/off LSTM-AE model outperformed the conventional LSTM-AE model.
- Subjects
SEOUL (Korea); SOUTH Korea; AIR compressors; SUBWAYS
- Publication
Journal of Mechanical Science & Technology, 2023, Vol 37, Issue 12, p6287
- ISSN
1738-494X
- Publication type
Article
- DOI
10.1007/s12206-023-1106-1