In diagnostic applications, data acquired from a unit in operation is often compared to predictions generated from a reference model. The equipment condition is often assessed via residual analysis, which compares the running data to model predictions. Often, the reference model may take the form of a high-fidelity, first principles physics model. Here, we seek to capture the dominant features of the turbine engine, using parameters typically instrumented in field applications using reduced rank linear models that are trained on data generated from the high fidelity design models. The reduced rank linear models are well suited to diagnostic applications. Specifically, a modified Principle Component Regression is applied to the reference data to obtain our reduced rank model. We then use real measurement data input into the reduced rank model to produce predictions and correspondingly, residuals whose statistical properties can characterized with respect to the high-fidelity model. This requires characterization of the expected measurement errors which are user input. The model is capable of working with a complete or reduced set of measurements. In the case of the redundant measurements, we can perform an analysis on the fidelity of the measurements. We show how to calculate a measure of agreement between a given set of measurements and the underlying model. Departures of the given, real data from the models predictions indicate possible faults in the operating variables. In addition, if a certain sensor is suspected of being fouled, we can leverage the ability of the model to predict with a reduced set of inputs and then compare the new predictions measure of agreement with the underlying model.

This content is only available via PDF.
You do not currently have access to this content.