Novelty detection methods for online health monitoring and post data analysis of turbopumps

As novelty detection works when only normal data are available, it is of considerable promise for health monitoring in cases lacking fault samples and prior knowledge. We present two novelty detection methods for health monitoring of turbopumps in large-scale liquid-propellant rocket engines. The first method is the adaptive Gaussian threshold model. This method is designed to monitor the vibration of the turbopumps online because it has minimal computational complexity and is easy for implementation in real time. The second method is the one-class support vector machine (OCSVM) which is developed for post analysis of historical vibration signals. Via post analysis the method not only confirms the online monitoring results but also provides diagnostic results so that faults from sensors are separated from those actually from the turbopumps. Both of these two methods are validated to be efficient for health monitoring of the turbopumps.