Application of Advanced Sampling Techniques to Handle Class Imbalance in GT Blade Failure Prediction

Abstract

Abstract The Blades of a Gas Turbine Engine are subjected to an exceedingly high temperature, pressure, and centrifugal load during its operation. Consequently, they exhibit a heightened susceptibility to failure modes such as creep, crack initiation and growth, HCF, Flutter, and Oxidation. Attaining accuracy in the prediction of these failure modes is of paramount importance. Traditionally this has been achieved by Finite element packages, discretizing the problem into finite elements, applying the necessary boundary conditions, and solving a mathematical equation for each of these finite elements. The process involves a numerical solution of differential equations that govern these physical phenomena. However, this pursuit is not without tradeoff, as it demands a significant investment of temporal, human, and computational resources. To minimize mentioned efforts, advanced data-driven techniques can be implemented to predict the creep failure instances. Machine learning is gaining significant traction for solving these kinds of problems because of advances in storage and computational power. Machine learning focuses on using data and algorithms to learn the pattern and make prediction on unseen data with substantial accuracy. Generally, machine learning can be classified into supervised and unsupervised learning. In this work we are developing a machine learning based classification model (supervised learning) to capture the failure instances. Failure prediction using Machine learning is generally encountered with higher prediction error because of imbalanced dataset (minority class is the focus of attention — less than 5% of data). This paper deals with the approach of improving turbine failure prediction using various sampling techniques. Random Oversampling, Synthetic Minority Oversampling Technique – SMOTE, Tomek Links and combine SMOTE-Tomek links sampling methods are utilized to generate a balanced dataset for improving Machine Learning model performance. Coolant mass flow rate, Hot gas mass flow rate, Coolant Temperature, Turbine Inlet Temperature and Blade Material are provided as an input to the Machine Learning model, and on the other hand, a binary variable indicating creep failure instances is set to be the target variable for the model. The Machine Learning models used for prediction are Naïve Bayes classifier, Random Forest, Adaboost and gradient boost methods. The results demonstrate that among all the sampling technique mentioned above, combined SMOTE-Tomek link sampling method resulted in maximum improvement of recall and F1 score.