Performance based creep life estimation for gas turbines application

Accurate and reliable component life prediction is crucial to ensure both the safety and economics of gas turbine operations. In the pursuit of improved accuracy and reliability, current model-based creep life estimation methods have become more and more complicated and therefore demand huge amounts of work and significant amounts of computational time. Because of the underlying problems arising from current life estimation methods, this research aims to develop an alternative performance-based creep life estimation method that is able to provide a quick solution to creep life prediction while at the same time maintaining the achieved accuracy and reliability as that of the model-based method. Using an artificial neural network, the existing creep life prediction subprocesses and secondary inputs are ‘absorbed’ into simple parallel computing units that are able to create direct mapping between various gas turbine operating and health conditions or gas path sensors and creep life. The outcome of this research is the creation of three proposed neural-based creep life prediction architectures known as the Range-Based, Functional-Based and Sensor-Based. An integrated creep life estimation model was first developed and incorporated into an in-house performance simulation and diagnostics software. Using the integrated model, the effects of several operating and health parameters on a selected turbo-shaft engine model turbine blade’s creep life was initially performed using an introduced Creep Factor approach. The outcomes of this investigation were then used to populate input-output samples to train and validate the neural-based creep life prediction architectures. To ensure that the proposed neural architectures are able to achieve generalisation and produce accurate creep life prediction for both clean and degraded engine conditions, four-stage assessments were carried out. Finally, the effects of input uncertainties on the creep life prediction were investigated to assess how sensitive the proposed architectures are to different levels of uncertainty. The results show that all of the proposed neural architectures were able to produce accurate creep life predictions for both clean and degraded engine conditions. When comparing the three proposed architectures, the Sensor-Based architecture was found to be the most accurate in both conditions. Despite the accurate creep life prediction, it was also found that all of the proposed architectures were sensitive to input uncertainties with the Functional-Based architecture being the least sensitive to the uncertainty.