Abstract
To address challenges in extracting health indicator (HI) curves and making accurate predictions with limited datasets in mechanical system prognostics, this study proposes a digital twin (DT)-driven framework for estimating remaining useful life (RUL). To minimize the deviation between simulated and measured data, we introduce a finite element model correction method using a stacked autoencoder-long short-term memory (SAE-LSTM) network. To reduce reliance on manual expertise and prior knowledge, the LSTM network is used to directly extract features from the frequency-domain vibration data and construct initial HI curves representing equipment performance degradation. Finally, this study employs a relevance vector machine (RVM) model to predict the HI curve trend by integrating failure criteria with twin data to establish the failure threshold. Experimental validation using the PHM2012 public dataset showed that the DT-based RUL prediction reduces the average relative error by 5.4% compared with traditional RUL prediction methods.