Predictive maintenance is an important application area for data-driven systems, especially when unexpected machine failures can cause downtime, repair costs, and safety-related risks. A central task in this area is Remaining Useful Life (RUL) prediction, which estimates how long a system is expected to operate before failure. This task is challenging because degradation is usually observed through multivariate sensor signals, and complete run-to-failure data may be limited in practical settings.

This thesis studies RUL prediction for turbofan engines using the NASA C-MAPSS degradation dataset. The aim is to compare selected feature-engineered machine learning models with sequence-based deep learning models under a controlled experimental setup. Random Forest, XGBoost, and Support Vector Regression are evaluated using statistical features created from sensor windows. LSTM, GRU, and 1D-CNN are evaluated using ordered sensor sequences. The study also investigates how prediction performance changes when the available training data is reduced.

The results show that recurrent deep learning models provide the strongest overall performance in the tested setup. In the direct 30-cycle comparison, GRU achieved the lowest MAE, while LSTM achieved the highest R². In the deep learning window-size experiment, LSTM with a 50-cycle window achieved the lowest MAE. The reduced data results show that LSTM remained the strongest model across all tested training data sizes, although performance varied between different selected engine subsets.

The thesis contributes a controlled empirical comparison of RUL prediction approaches and shows how model type, sequence length, and training data availability influence prediction performance.