Neural Networks for the Diagnostics of Gas Turbine Engines

Abstract

The paper describes the activities carried out for developing and testing Back Propagation Neural Networks (BPNN) for the gas turbine engine diagnostics. One of the aims of this study was to analyze the problems encountered during training using large number of patterns. Each pattern contains information about the engine thermodynamic behaviour when there is a fault in progress. Moreover the research studied different architectures of BPNN for testing their capability to recognize patterns even when information is noised. The results showed that it is possible to set-up and optimize suitable and robust Neural Networks useful for gas turbine diagnostics. The methods of Gas Path Analysis furnish the necessary data and information about engine behaviour. The best architecture, among the ones studied, is formed by 13, 26 and 47 neurons in the input, hidden and output layer respectively. The investigated Nets have shown that the best encoding of faults is the one using a unitary diagonal matrix. Moreover the calculation have identified suitable laws of learning rate factor (LRF) for improving the learning rate. Finally the authors used two different computers. The first one has a classical architecture (sequential, vectorial and parallel). The second one is the Neural Computer, SYNAPSE-1, developed by Siemens.

Engineering Creative Design in Robotics and Mechatronics

“…They employed genetic programming tool for NARMAX and B-spline models to determine the model structure. Different configurations of Back Propagation Neural Networks (BPNN) were used by Torella, Gamma and Palmesano to study and simulate the effects of gas turbine air system on engine performance (Torella et al, 2003).…”

Performance-analysis-based gas turbine diagnostics: A review

“…Application of Feed-Forward Back-Propagation neural networks to gas turbine diagnosis has been performed by many researchers, such as Denney [ [76] and so on. Torella and Lombardo [70] described a calculation for the learning rate factor (LRF), for improving the learning rate of back propagation neural networks (BPNN). Kanelopoulos et al [71] presented a partial network architecture to perform sensor and component fault diagnosis step by step.…”

Recurrent Neural Network Based Simulation of a Single Shaft

“…ML-based techniques like ANN have shown the capability to predict dynamic behavior of GTs without having access to information about the system physics. Different ANN-based methodologies have already been investigated and developed in order to disclose complex nonlinear behavior of aero gas turbines (Agrawal and Yunis, 1982;Chiras et al, 2001Ruano et al, 2003;Torella et al, 2003;Sarkar et al, 2012Sarkar et al, , 2013Salehi and Montazeri, 2018;Ibrahem et al 2019). These efforts have covered a variety of approaches such as…”