BROWSE
Detail
Journal Article
Contrastive representation learning for surge prediction in air compressor systems
Cited 0 time in 
Share 
Share
- Authors
- Nakyoung Kim, Jangkyum Kim, Yoon-Sik Yoo, Il-Woo Lee
- Issue Date
- 2026-01
- Citation
- Measurement: Journal of the International Measurement Confederation, v.258, no.Part C, pp.1-13
- ISSN
- 0263-2241
- Publisher
- Elsevier
- Language
- English
- Type
- Journal Article
- Abstract
- Surge is a dynamic instability phenomenon observed in air compressor systems, which results in operational inefficiencies and monetary loss for industrial operations. Surge in compressor systems is considered critical as it is associated with severe and abrupt abnormal airflow dynamics and, thereby, immediate corrective actions are required, such as excessive air release. Accordingly, there is a growing need for proactive management enabled by the digital transformation of the systems. To this end, this study proposes a domain-specific contrastive representation learning mechanism with deep neural networks to predict surge events before they fully develop. The effectiveness and efficiency of the proposed model are experimentally demonstrated using real-world data collected from an operational air compressor installed on an industrial site. Evaluation with the real-world dataset demonstrates the proposed model's capability to predict surge events a minute in advance of their occurrence, with accuracy higher than 95% and an F
1 score above 92%. Furthermore, the experimental results show that the proposed model outperforms conventional Machine Learning (ML) baselines in terms of prediction accuracy. The proposed model also achieves comparable or superior prediction performance to complex Deep Learning (DL) approaches while significantly reducing model complexity in terms of computational operations, inference time, and model size. These results underscore the proposed model's practicality and scalability for real-world industrial environments, especially in systems operating with continuous streaming sensor data.
- KSP Keywords
- Air compressor, Air release, Corrective actions, Data collected, Deep neural network(DNN), Domain-specific, Dynamic instability, Effectiveness and efficiency, Industrial environment, Industrial site, Machine learning (ml)