Uncertainty Prediction of Remaining Useful Life Using Long Short-Term Memory Network Based on Bootstrap Method

The accuracy and reliability of remaining useful life (RUL) prediction have a significant influence on Condition Based Maintenance (CBM). Deep learning methods, especially Long Short-Term Memory networks (LSTM), have been proven to be effective in automatically mining the hidden features of sensors data, and then obtain a better point estimation for RUL forecast. However, it is pointless when the sensor data is full of noise. Estimated RUL often varies widely mainly due to the model parameters and the noise. In this case, it is inadequate to calculate the point estimation of RUL merely. For uncertainty prediction of RUL, this requires calculating not only the deterministic prediction value but also the estimation of the prediction interval. The reliability of deterministic RUL prediction, incorporated into the uncertainty prediction, is enhanced. In this paper, an LSTMFNN (Feedforward Neural Network) based on bootstrap method (LSTMBS) for uncertainty prediction of RUL estimation is presented. The proposed method is tested on the benchmarking aircraft turbofan engines datasets (C-MAPSS). Different feature augmentation methods are applied to different sub-datasets of C-MAPSS due to the different operating conditions and fault modes. Results show that the proposed approach not only performs a competitive point estimation of RUL compared with other approaches but also obtains additional information on prediction interval of RUL to determine the appropriate maintenance time. In order to address this problem more universally, the unified LSTMBS model is constructed considering not discriminating the categories of new test trajectories in the practical system.