Remaining Useful Life Prediction Based on Bayesian Inference Long and Short-Term Memory Networks

Remaining useful life prediction using deep learning methods has been widely studied, however most deep learning methods ignore the widespread problem of prediction uncertainty in equipment, which has an impact on health prediction and subsequent decision making. Less fault data leads to data imbalance and model uncertainty in prediction, which has an impact on model training. A new Bayesian Inference Long and Short-Term Memory Networks (Bayes-LSTM) is constructed for Remaining useful life prediction. Firstly, the data set is expanded to reduce data imbalance by generating fake data using generative adversarial networks, and secondly, the parameters of the long and short-term memory network model are represented as random variables in a Bayesian framework, and the posterior distribution is approximated using variational inference for model training, which leads to the probability distribution of RUL prediction. Finally, the performance of the proposed method is evaluated through experimental analysis on the CMAPSS dataset, while 95% confidence intervals are used to quantify the uncertainty. Experiments show that the network is effective, it can well reduce the uncertainty of the model, effectively improve the performance degradation trend prediction ability and improve the prediction accuracy.