Predicting the Remaining Useful Life of a Gas Turbine Based on an Exponential Degradation Model

In the context of rotating machinery maintenance, this study examines diagnostic, prognostic and health management (PHM) techniques, where a set of signal processing and supervised machine learning processes are used in a precise strategy that allows us to diagnose the condition of a gas turbine and also to predict its remaining useful life (RUL), to make appropriate maintenance decisions. Using a database of historical vibration signals, partial failure due to misalignment is diagnosed by Fast Fourier Transform (FFT) analysis. This partial failure can evolve into a total failure, leading to the cessation of production and higher maintenance costs. Predictive maintenance can improve its ability to predict failure phenomena using data-driven degradation modeling. On this basis, a total failure of a gas turbine can be avoided. A set of temporal and frequency health indicators are extracted at each operating cycle to be analyzed, filtered, and improved using statistical techniques such as normalization methods and correlation analysis. Additionally, the rolling average is used based on window functions and monotonicity to choose the best and most reliable indicators. Before fitting the decay curves of the selected health indicators using an exponential regression model, they were combined into a single indicator by applying the principal component analysis (PCA) technique to facilitate the adjustment process. The results showed good performance after testing the exponential model to predict the RUL on the test data.