Koopman theory assisted transfer learning for anomaly detection of liquid rocket engines in frequency domain

As the foundation for various fault diagnosis work, Anomaly Detection plays an important role in ensuring the safety and reliability of various complex systems, Liquid Rocket Engine, for example. However, real-world signals exhibit intrinsic non-stationary characteristics, due to the inherent noise present in actual work situations, resulting in a significant challenge for time-series based anomaly detection. On the other hand, the scarcity of actual working data for Liquid Rocket Engine also brings certain challenges. In this study, a novel Furior-Koopman based neural Network (FKN) is proposed for time-series anomaly detection of Liquid Rocket Engine. The proposed Furior-Koopman models time-series data in the frequency domain via Koopman operator, which has significant advantages in reducing noise interference. A transfer learning method from simulated data to actual data is also applied to handle the problem of insufficient actual data quantity. In order to meet the demand for anomaly detection of liquid rocket engine, we have established a dataset for anomaly detection of rocket engines and several experiments are conducted on established dataset. The experimental results demonstrate that the proposed method outperforms several recent methods. Several experiments are also conducted on common datasets to show the effectiveness of proposed method. All data and code realted can be found at https://github.com/minussix/FKN_for_LRE .