Bidirectional Long Short-Term Memory Network Bayesian Optimization for Aero-Engine Fault Prediction

In this paper, a fault prediction method for aero-engines is proposed based on a bidirectional long short-term memory (BiLSTM) network. First, a BiLSTM network is constructed for analyzing multi-variable sequential sensor data. The output of the BiLSTM network incorporates information from both the forward and backward temporal sequences processed by the LSTM network. The long-term dependency issue of recurrent neural networks (RNNs) is effectively addressed by the LSTM, which retains essential input information. Secondly, the parameters of the BiLSTM network are optimized by using the Bayesian optimization algorithm, which reduces training time and achieves superior results compared to manual parameter tuning method. Thirdly, aero-engine failures are predicted by using the BiLSTM method. Finally, experimental results on the aero-engine uncertain performance degradation fault datasets demonstrate that the proposed model outperforms RNN and LSTM prediction methods.