On-Line Validation of Measurements on Jet Engines Using Automatic Learning Methods

Nowadays, turbine engine tests are processed using an open loop, i.e. the measurements are verified and treated a posteriori, sometimes weeks or months after the end of the test. The purpose of the present project is to develop a new methodology which enables real time detection of faulty measurements and the suppression of the source of these faults during the test.
The validation of the measurements is achieved by a “robust” parameter identification [1]. Such a
method is called robust in the sense that it can cope with 20 to 30% of faulty measurements. The robustness is insured by a distribution of the measurement noise, as introduced by Huber [7, 8], that takes into account the possibility of faults. The purpose of a parameter identification is to find the set of parameters which has most likely generated the measurements observed on the process.
This leads to an optimisation problem that has to be solved for the parameters. The measurements are linked to the parameters through a non-linear model, leading to a large system of equations for modern jet engines. If no physical model of the process can be made available or if this model is too complex to allow real time validation, automatic learning methods may provide a solution:
• either a mathematical representation is generated, directly based on the measurements (online
learning),
• or a database is first generated, based on the existing (but expensive) physical model, the
database being subsequently used to build a statistical model (off-line learning).
Neural networks seem to be very suitable for modeling the behavior of turbojets, avoiding the resolution of a time-consuming non-linear system. In this paper neural networks are tested to generate a mathematical representation of a single flow, single spool and variable geometry nozzle turbojet, from a data base of “measurements” generated by a physical model of the engine. Only the off-line learning approach is considered.