Remaining useful life prediction by degradation distribution transport health indicator and consolidated memory stabilized LSTM

Accurate prediction of the remaining useful life (RUL) is of paramount importance for preventing unexpected failures in industrial machinery. This primarily involves the construction of health indicator (HI) to capture degradation information and the establishment of relationships between HI and RUL. However, most previous methods which require complex model structures or rich domain knowledge are not suitable for real industrial conditions with variable operating conditions. To address this challenge, a novel RUL prediction framework is proposed based on a degradation distribution transport health indicator (DDTHI) and a consolidated memory stabilized LSTM (CMsLSTM). First, a new degradation data distribution transport matrix is proposed, requiring no prior domain knowledge, to characterize the transformation process between degradation data distributions. Then, the HI at the current degradation time is constructed by minimizing the distribution transport cost. To streamline the prognostic architecture while preserving degradation information retention capability, a consolidated memory stabilized LSTM is designed via optimizing the internal structure of neurons and relearning the relationship between different time points. It can further store historical knowledge across time scales and provide more information for the RUL prediction. Finally, the effectiveness of the proposed method is verified by a real motor bearing dataset of a combustion fan in a hot strip mill and a public bearing dataset.