Uncertainty Quantification For Remaining Useful Lifetime Prediction With Multi-Channel Sensory Data

For remaining useful lifetime (RUL) prediction with multi-channel sensory data, long-term prediction has more uncertainty than short-term prediction. In this paper, the ratio of mean to variance was considered to measure the uncertainty propagation rate (UPR) of RUL prediction over time. Furthermore, we use a recurrent neural network (RNN) as the linking function for the mean of inverse Gaussian distributed RUL to construct a two-stage hybrid model. Later the RNN and the UPR are jointly trained with sensory data and failure records via alternating minimization. Proposed algorithms are validated in a simulation test.