Application of deep neural networks method to the fault diagnosis during the startup transient of a liquid-propellant rocket engine

Abstract

Methods of deep neural networks (DNN) such as convolutional neural network (CNN) and long short-term memory (LSTM) have been proposed as new solutions for the fault detection and identification (FDI) problems. The methods are robustness to the noise, capable of approximate high-dimensional nonlinearity, and cost-effective. In this study, we introduced a DNN method for FDI of a liquid-propellant rocket engine (LRE) during its early phase of startup transient. A numerical simulator that can describe the normal and abnormal startup transient of the LRE was developed and validated through comparison with actual hot-firing test data. A dataset representing 33 classes of faults during the startup transient of an open-cycle LRE was created using the simulator to train the DNN. The FDI problem is divided into two steps: fault detection step and fault classification step. The fault detection step provides the binary determination for the condition (normal or abnormal) of the LRE system. Then the fault classification step determines the fault type to which the current abnormality belongs. Each FDI step uses an LSTM forecasting model to take advantage of its ability to grasp temporal contexts for time series data. A CNNLSTM classification model was implemented by stacking the time series data into a two-dimensional image to extract the spatial correlations for the classification problem. A multivariate statistic (the Mahalanobis distance) is used to determine the normality of the experiment. The effectiveness of the proposed framework – particularly in comparison with the conventional red-line method – was demonstrated through a case study.