Development of Machine Learning Model to Predict Mechanical Failure of Gas Turbine Disc

Abstract

Abstract Gas Turbine is a mechanical system which is used for power generation since decades. Its components are highly stressed and exposed to very high temperature. It operates on various environmental condition which experience numerous loading patterns. In all these operating conditions, Gas turbines must operate with higher efficiency to meet changing power requirements. Each component must be assessed for different mechanical failures and predict life during design and development. Gas Turbine Rotors which comprise of several discs are one of the most critical components in gas turbine. GT rotor is subjected to high temperature and centrifugal load during their operations. In such conditions, predicting the failures becomes utmost priority. This study aimed to evaluate the effectiveness of Machine Learning (ML) techniques to predict the failure in a gas turbine Rotor Disc. For this study, a simplified geometrical 3D model of Turbine Disc was used along with Cooling holes, Hub geometry with Sealing arm and lifting features. Four input variables i.e., cooling hole mass flow, temperature of cooling air, purge hole mass flow and temperatures of purge air were used as input feature to the machine learning model. Steady-state thermo-mechanical analysis were performed to evaluate the metal temperature and subsequently the stress and life of the component under various load cases. A machine learning based surrogate model was developed based on the data extracted from 3D thermo-mechanical FEA assessment. The generated dataset was randomly divided into 75:25 ratio for training and testing of ML models respectively. Multiple models based on different algorithms were created for predicting disc LCF life. Then, these were evaluated with test set to select the model using various Evaluation Metric. Machine Learning techniques such as Logistic Regression, Random Forest and Support Vector Machine algorithms were compared using Precision, Recall and F1 Score. The results were validated through Confusion Matrix and ROC (Receiver Operating Characteristics) curves. This study demonstrates that ML techniques has potential in predicting failures/life of a component. This study will be helpful for the assessments performed during later stage of product life cycle for example overhaul, lifetime extension or if any manufacturing deviations happen. Mostly these assessments are very critical and time dependent. As we use conventional methods for component assessments, it takes significant time and cost. Therefore, this study and its implementation would make current industrial practices efficient and cost effective.