Fault Detection in Liquid Rocket Engines via Segmented Vector Autoregressive Modeling

Abstract

Ensuring the safe operation of liquid rocket engine (LRE) systems requires reliable fault diagnosis, yet the scarcity of real fault data limits deep learning applications despite their modeling strengths. We address this by developing an offline detection method based on piecewise stationary Vector Autoregressive (VAR) modeling, employing a two-phase approach that first identifies candidate change points through block fused LASSO regularization and subsequently refines them using Smoothly Clipped Absolute Deviation (SCAD) regularization to leverage its asymptotic unbiasedness. Validated on a high-fidelity LRE simulation dataset (26 sensors, 2000 time points) with injected faults including turbopump efficiency degradation, hydrogen turbine leakage, and valve failures across 48 scenarios, our method achieves 100% precision (±50-sample tolerance) in fault timing detection without requiring training data, demonstrating superior performance to conventional ARMA models while overcoming the data dependency of neural networks.