Gas turbine blades and vanes face very severe operating conditions - high temperature and pressure which necessitates the creation of complex cooling and component designs, resulting in high computational cost. The ability to predict cyclic failure in these components is therefore a critical activity that has been historically performed using 3D commercial finite element (FE) codes for baseload conditions. However, these codes take substantial time and resources which restricts their application in failure prediction at variable operating conditions. Newer data-driven techniques such as machine learning (ML) provide a valuable tool that can be utilized to predict the occurrence of cyclic failure for these conditions with minimal time and resource requirement.
In this paper, a machine learning based surrogate model is developed to predict the cyclic failure of a radially cooled turbine blade. The features used as input to machine learning model are turbine inlet temperature, coolant inlet temperature, hot gas mass flow rate, cooling air mass flow rate and blade materials. The output for the model is a binary variable depicting the incident of component failure. 70% of the FE data points are used to train the ML model while the remaining are used for testing. A comparative study between Logistic Regression, Random Forest, K-nearest neighbor, and Support Vector Machine (SVM) was performed to select the most accurate algorithm for the classification model. Finally, the results show that the Random Forest and SVM algorithms predicts failure with the highest f-1 score of 0.92. The model also demonstrates that Turbine Inlet temperature has the highest importance amongst the input features followed by blade material. Additionally, this methodology offers a tremendous advantage for failure prediction by reducing analysis time from multiple hours to a few seconds, rendering this technique especially beneficial for time sensitive business decisions in the gas turbine industry.