Model-based supervised sensor placement optimization to detect propellant leak in a liquid rocket engine

The sensor selection for leak detection and diagnosis in a reusable liquid rocket engine was studied. Sufficient amounts of normal and leaked training data were generated by Monte Carlo simulations considering variations in system conditions based on past firing test results. A multivariate supervised analysis, in which measurements are linearly projected onto a vector that characterizes the difference between distributions of normal and leaked cases, successfully detected the simulated leaks that could not be detected by conventional univariate red-line judgment and unsupervised principal component analysis. Significant sensors for leak detection were selected using the greedy approach based on the detection performance score. It was found that only 11 of the 27 sensors were sufficient to maintain the detection performance.