Multi-Scale Remaining Useful Life Prediction Using Long Short-Term Memory

Predictive maintenance based on performance degradation is a crucial way to reduce costs and potential failures in modern complex engineering systems. Reliable remaining useful life (RUL) prediction the main criterion for decision-making predictive maintenance. Conventional model-based methods data-driven approaches often fail achieve an accurate result using single model system featuring multiple components operational conditions, as pattern usually nonlinear time-varying. This paper proposes novel multi-scale RUL approach adopting Long Short-Term Memory (LSTM) neural network. In feature phase, Pearson’s correlation coefficient applied extract representative features, operation-based data normalisation presented deal with cases where patterns are concealed sensor data. Then, three-stage target function proposed, which segments process of into non-degradation stage, transition linear stage. The classification these three stages regarded small-scale prediction, it achieved through processing signals after LSTM-based binary algorithm combined method. After that, specific built last two produce large-scale prediction. proposed validated by comparing several state-of-the-art techniques widely used C-MAPSS dataset. A significant improvement most subsets. For instance, 40% reduction Root Mean Square Error over best existing method subset FD001. Another contribution that offers more degree flexibility strategy depending availability stage in.