Neural Network Approach for Predicting the Failure of Turbine Components

Turbine components operate under severe loading conditions and at high and varying temperatures that result in thermal stresses in the presence of temperature gradients created by hot gases and cooling air. Moreover, static and cyclic loads as well as the motion of rotating components create mechanical stresses. The combined effect of complex thermo-mechanical stresses promote nucleation and propagation of cracks that give rise to fatigue and creep failure of the turbine components. Therefore, the relationship between thermo-mechanical stresses, chemical composition, heat treatment, resulting microstructure, operating temperature, material damage, and potential failure modes, i.e. fatigue and/or creep, needs to be well understood and studied. Artificial neural networks are promising candidate tools for such studies. They are fast, flexible, efficient, and accurate tools to model highly non-linear multi-dimensional relationships and reduce the need for experimental work and time-consuming regression analysis. Therefore, separate neural network models for γ’ precipitate strengthened Ni based superalloys have been developed for predicting the γ’ precipitate size, thermal expansion coefficient, fatigue life, and hysteresis energy. The accumulated fatigue damage is then estimated as the product of hysteresis energy and fatigue life. The models for γ’ precipitate size, thermal expansion coefficient, and hysteresis energy converge very well and match experimental data accurately. The fatigue life proved to be the most challenging aspect to predict, and fracture mechanics proved to potentially be a necessary supplement to neural networks. The model for fatigue life converges well, but relatively large errors are observed partly due to the generally large statistical variations inherent to fatigue life. The deformation mechanism map for 1.23Cr-1.2Mo-0.26V rotor steel has been constructed using dislocation glide, grain boundary sliding, and power law creep rate equations. The constructed map is verified with experimental data points and neural network results. Although the existing set of experimental data points for neural network modeling is limited, there is an excellent match with boundaries constructed using rate equations which validates the deformation mechanism map