Predicting engine parameters using the optical spectrum of the Space Shuttle Main Engine exhaust plume

Ashok N. Srivastava, Wray Buntine1995 · DOI

TL;DR

This paper presents a novel approach to predict Space Shuttle Main Engine (SSME) parameters like Rated Power Level (RPL) by analyzing its exhaust plume's optical spectrum. They developed a method to decompose the high-dimensional spectral data into meaningful components, which are then fed into a neural network for accurate RPL prediction.

What Problem Does This Paper Address?

Monitoring the health and performance of complex rocket engines like the SSME is crucial for safety and efficiency. The Optical Plume Anomaly Detection (OPAD) system aims to do this by analyzing the exhaust plume's optical spectrum. While detecting anomalies based on metal concentrations is one goal, this paper focuses on a more challenging task: predicting key engine parameters such as the Rated Power Level (RPL) and Mixture Ratio (MR) directly from the spectrum. The primary challenge is the extremely high dimensionality of the spectral data (thousands of wavelength points), making direct analysis computationally prohibitive and prone to poor prediction results.

Methods

The authors tackle the high dimensionality by first decomposing the optical spectrum into a set of orthogonal components. They propose a linear algorithm to achieve this decomposition, which they claim yields results similar to Singular Value Decomposition (SVD) but is more computationally feasible for their data. This decomposition transforms the raw, high-dimensional spectral data into a lower-dimensional set of "features" (coefficients of the components).

These extracted features are then used as input to a multilayer feedforward neural network. The network is designed to learn the inverse mapping from the spectral features to the engine parameters. Specifically, they focus on predicting the RPL, which is characterized by piecewise constant behavior. The neural network architecture includes an input layer, a hidden layer with 8 tanh units, and a single output unit. They use the backpropagation algorithm for training and minimize the sum-squared error between the predicted and actual RPL.

Datasets

The paper uses data from approximately 30 tests of the SSME, where each test consists of spectral measurements taken at 0.5-second intervals for about 200 seconds. The spectral data covers wavelengths from 3000-9500 Angstroms, resulting in data matrices of size Nxp, where N is the number of time steps and p is the number of wavelength points (around 4500). The corresponding ground truth for engine parameters like RPL and MR are also provided for these tests.

Key Insights and Ideas

  • Spectral Decomposition for Feature Extraction: The core idea is that the complex optical spectrum can be reduced to a few physically interpretable components. This significantly reduces dimensionality and makes the data more amenable to machine learning.
  • Orthogonal Components: The algorithm's ability to produce orthogonal components is highlighted as a way to ensure that each component captures unique information, potentially leading to better predictive models.
  • Neural Networks for Inverse Mapping: Recognizing that the relationship between the spectrum and engine parameters is complex and likely nonlinear, neural networks are employed as powerful function approximators to learn this inverse mapping.
  • Piecewise Constant Nature of RPL: The authors exploit the observation that RPL is a piecewise constant signal, which influences their choice of a sigmoidal output unit in the neural network, treating it as a classification problem.

Key Results

The paper demonstrates that the neural network, using the extracted spectral features, can predict the Rated Power Level (RPL) with good accuracy. Figure 6 shows that the neural network's predictions closely follow the actual RPL signal for both training and out-of-sample test sets. The paper quantifies the performance by comparing the sum-squared errors of the neural network predictor against a simple linear predictor, showing a significant improvement (ratios of SSElin/SSEnet ranging from 1.034 to 7.279). The neural network is shown to be more robust and better at capturing the trend of the RPL signal compared to the linear predictor.

Limitations and Future Work

  • Focus on RPL: The current work primarily focuses on predicting the Rated Power Level (RPL). The prediction of the Mixture Ratio (MR) is acknowledged as a more challenging task, expected to be more involved due to its non-piecewise constant nature.
  • Orthogonal Feature Vectors: While the algorithm produces orthogonal component vectors, future research could explore creating orthogonal feature vectors, which might lead to even better predictions.
  • Comparison with Other Predictors: The authors suggest comparing their neural network approach with other, more complex linear predictors.
  • Real-time Implementation: The neural network predictors for RPL have been implemented in a real-time system at the Marshall Space Flight Center, but results from this implementation were not yet available at the time of publication.
  • Computational Cost: While their decomposition method is more feasible than direct PCA on the high-dimensional data, the computational aspects of processing such large spectral datasets remain a consideration.

Summary updated

Abstract

The Optical Plume Anomaly Detection (OPAD) system is under development to predict engine anomalies and engine parameters of the Space Shuttle's Main Engine (SSME). The anomaly detection is based on abnormal metal concentrations in the optical spectrum of the rocket plume. Such abnormalities could be indicative of engine corrosion or other malfunctions. Here, we focus on the second task of the OPAD system, namely the prediction of engine parameters such as rated power level (RPL) and mixture ratio (MR). Because of the high dimensionality of the spectrum, we developed a linear algorithm to resolve the optical spectrum of the exhaust plume into a number of separate components, each with a different physical interpretation. These components are used to predict the metal concentrations and engine parameters for online support of ground-level testing of the SSME. Currently, these predictions are labor intensive and cannot be done online. We predict RPL using neural networks and give preliminary results.
BibTeX
@misc{srivastava1995predicting,
  author = {Ashok N. Srivastava and Wray Buntine},
  title = {{Predicting engine parameters using the optical spectrum of the Space Shuttle Main Engine exhaust plume}},
  year = {1995},
  doi = {10.2514/6.1995-954},
}
Figure