A Shape-constrained Transfer Temporal Transformer Network for remaining useful life prediction of rotating machines

Predicting remaining useful life (RUL) of rotating machines can promote predictive maintenance and improve the mechanical system reliability. This paper proposes a Shape-constrained Transfer Temporal Transformer Network (SC3TN) for RUL prediction across different rotating machines. The disentangled representation learning is utilized to extract invariant features across different domains. Health indexes are constructed with invariant features considering shape constraints and all previous health states. The operating time is encoded by a Temporal Encoding in predicting process. The SC3TN is tested on two tasks, namely RUL prediction across different working conditions and different equipment. Compared to state-of-the-art methods, SC3TN reduces the mean absolute error by 4.94% and 5.47% and improves the score by 8.5% and 20.83% on two tasks. The experiments show that the proposed method can achieve more accurate RUL prediction than existing methods, which demonstrates the potential of the proposed method for predicting RUL without collecting degradation data in advance.