TY - GEN
T1 - NONLINEAR SURROGATE MODELS FOR GAS TURBINE DIAGNOSIS
AU - Loboda, Igor
AU - Castillo, Iván González
AU - Yepifanov, Sergiy
AU - Zelenskyi, Roman
N1 - Publisher Copyright:
Copyright © 2022 by ASME.
PY - 2022
Y1 - 2022
N2 - In gas turbine diagnostics, a significant contradiction is observed between the variety of methods proposed and a limited number of the algorithms realized in real monitoring systems. One of the explanations is the shortage of simple, but reliable and accurate diagnostic solutions. Many actual solutions are based on a nonlinear thermodynamic model; however, this physics-based model and an iterative procedure to adapt it are complex, critical to computer resources, and does not always converge. On the other hand, the used linear models are simple, but not accurate enough. The present paper proposes and analyzes two types of nonlinear simplified static data-driven models based on the steady-state data generated by the thermodynamic model. The first type includes direct models that compute monitored variables Y for given operating conditions U and fault parameters δΘ. The second type presents inverse models that estimate parameters δΘ using variables U and Y as inputs. The paper aims to create such models, optimize them, and show that they can be a good surrogate for the original thermodynamic model and its adaption procedure. Based on the experience of the development of baseline models, we employ polynomials and multilayer perceptron as approximation techniques. Adjustment of the techniques and their comparison allow choosing the best one. To draw solid conclusions on the utility of the proposed models, three different engines are used as test cases. The results of the verification of these models are promising for their use in gas turbine diagnostic and control systems.
AB - In gas turbine diagnostics, a significant contradiction is observed between the variety of methods proposed and a limited number of the algorithms realized in real monitoring systems. One of the explanations is the shortage of simple, but reliable and accurate diagnostic solutions. Many actual solutions are based on a nonlinear thermodynamic model; however, this physics-based model and an iterative procedure to adapt it are complex, critical to computer resources, and does not always converge. On the other hand, the used linear models are simple, but not accurate enough. The present paper proposes and analyzes two types of nonlinear simplified static data-driven models based on the steady-state data generated by the thermodynamic model. The first type includes direct models that compute monitored variables Y for given operating conditions U and fault parameters δΘ. The second type presents inverse models that estimate parameters δΘ using variables U and Y as inputs. The paper aims to create such models, optimize them, and show that they can be a good surrogate for the original thermodynamic model and its adaption procedure. Based on the experience of the development of baseline models, we employ polynomials and multilayer perceptron as approximation techniques. Adjustment of the techniques and their comparison allow choosing the best one. To draw solid conclusions on the utility of the proposed models, three different engines are used as test cases. The results of the verification of these models are promising for their use in gas turbine diagnostic and control systems.
KW - Artificial neural networks
KW - Gas turbine diagnostics
KW - Polynomials
KW - Surrogate models
UR - http://www.scopus.com/inward/record.url?scp=85141353135&partnerID=8YFLogxK
U2 - 10.1115/GT2022-83550
DO - 10.1115/GT2022-83550
M3 - Contribución a la conferencia
AN - SCOPUS:85141353135
T3 - Proceedings of the ASME Turbo Expo
BT - Coal, Biomass, Hydrogen, and Alternative Fuels; Controls, Diagnostics, and Instrumentation; Steam Turbine
PB - American Society of Mechanical Engineers (ASME)
T2 - ASME Turbo Expo 2022: Turbomachinery Technical Conference and Exposition, GT 2022
Y2 - 13 June 2022 through 17 June 2022
ER -