Hierarchical RUL Prediction for Turbofan Engines Based on Health Stage Classification and Change Point-Guided Data Augmentation

In prognostics, it is essential to accurately predict the Remaining Useful Life (RUL) of complex systems like turbofan engines. Most studies adopt data-driven methods and rely on piecewise linear (PwL) labeling, where a fixed initial RUL is assigned to all engines. To better reflect real-world degradation, this assumption is refined using change point detection to assign engine-specific initial RUL values, offering a more realistic problem formulation. In line with PwL-based labeling approaches, where mechanical systems shift from a normal to a degradation phase at a specific change point, distinct RUL prediction models are developed for each health stage. A classification model is proposed using a novel training methodology to identify the current health stage of an engine. This method segments the dataset into pre- and post-change intervals and applies data augmentation both to simulate the truncated nature of the test data—where engines are monitored up to unknown failure points—and to construct a balanced training set that fairly represents both healthy and degraded conditions, thereby improving the robustness and generalization of the classification model. Two deep learning models are used for RUL prediction in engines: an LSTM model for degraded engines, which captures long-term dependencies in deterioration data, and a feature-level fusion model for healthy engines with sparse sensor data, which enhances prediction accuracy by integrating CNN and LSTM, addressing unique problems in different engine health stages. Model interpretability is enhanced through post-hoc SHAP analysis, revealing stage-specific sensor importance and explaining predictions across different health stages.