Publications

The critical in our oil and gas company as a failure can led to significant production and financial losses are gas turbines. Based on our operational perspective and within the scope of our study, we have experienced considerable production and financial impacts due to maintenance issues resulting from gas turbine compressor failures. There is a risk of repetitive failure or worsening failure to manage. Meanwhile, these machines have 100 monitored parameters stored in historical data for both engines. Previously, this data was managed manually with significant gaps of inaccuracy between the actual and predicted failures as referred to in the operator annually. There is potential to improve the correlation analysis and predictive modeling using machine learning.

This manuscript study aims to predict failure modes for Gas Turbines, specifically in engine 1 combustor cracks and engine 2 thrust balance damage cases, by modeling normal operating conditions with machine learning. By identifying deviations from expected behavior, the model can help prevent premature failures, unplanned shutdowns, and costly maintenance. The gas turbine engine is installed in different plants but has a relatively similar gap of failure prediction accuracy, to sample the modelling simulation.

Regarding the advanced algorithm architecture and relatively similar data to process, the three deep learning algorithms MLP, CNN, and LSTM were selected. These algorithms are trained with pre-failure turbine data and tested using five-fold cross-validation. Performance was evaluated with RMSE and R² scores. LSTM outperformed the others, achieving the lowest RMSE (0.0234 for Engine 1, 0.0539 for Engine 2) and highest R² (0.8883 and 0.7677, respectively). This model can detect deviations, indicating potential failures. This approach can be applied to enhance turbine reliability, reduce maintenance costs, and improve efficiency.