Gas Turbine and Sensor Fault Diagnosis With Nested Artificial Neural Networks

Abstract

Accurate gas turbine diagnosis relies on accurate measurements from sensors. Unfortunately, sensors are prone to degradation or failure during gas turbine operations. In this paper a stack of decentralised artificial neural networks are introduced and investigated as an approach to approximate the measurement of a failed sensor once it is detected. Such a system is embedded into a nested neural network system for gas turbine diagnosis. The whole neural network diagnostic system consists of a number of feedforward neural networks for engine component diagnosis, sensor fault detection and isolation; and a stack of decentralised neural networks for sensor fault recovery. The application of the decentralised neural networks for the recovery of any failed sensor has the advantage that the configuration of the nested neural network system for engine component diagnosis is relatively simple as the system does not take into account sensor failure. When a sensor fails, the biased measurement of the failed sensor is replaced with a recovered measurement approximated with the measurements of other healthy sensors. The developed approach has been applied to an engine similar to the industrial 2-shaft engine, GE LM2500+, whose performance and training samples are simulated with an aero-thermodynamic modelling tool — Cranfield University’s TURBOMATCH computer program. Analysis shows that the use of the stack of decentralised neural networks for sensor fault recovery can effectively recover the measurement of a failed sensor. Comparison between the performance of the diagnostic system with and without the decentralised neural networks shows that the sensor recovery can improve the performance of the neural network engine diagnostic system significantly when a sensor fault is present.