Remaining useful life prediction based on hybrid CNN-BiLSTM model with dual attention mechanism

Abstract

• Proposes a dual-attention parallel CNN-BiLSTM model for aircraft engine RUL prediction. • Fuses sensor signals into Health Indicators (HI) to enhance data trends and prediction accuracy. • Uses Efficient Channel Attention (ECA) to adaptively weight multi-sensor degradation features without dimensionality reduction. • Extracts temporal features via BiLSTM and multi-head attention for richer degradation sequence modeling. • Validated on aircraft engine data: significantly improves RUL prediction performance over baselines. The precise prediction of the remaining useful life (RUL) of aircraft engines holds significant importance for airlines in formulating optimal maintenance strategies and efficiently curbing maintenance expenses. CNN is used to extract spatial sequence features and LSTM is used to capture temporal sequence characteristics in the prediction approach for aviation engine RUL. However, in the mainstream approach, both CNN and LSTM are connected in a serial manner, resulting in significant information loss and redundant computation. We present a new parallel model in this research that includes a dual attention mechanism, leveraging both CNN and BiLSTM networks, to accurately forecast the RUL of aircraft engines. Firstly, The health index (HI) is created by fusing the preprocessed sensor signals, which serves as the input sequence along with the joint sensor signals. Subsequently, a parallel network structure comprising CNN and BiLSTM is formulated, integrating the channel attention (ECA) module and multi-head attention optimization techniques to extract spatial and temporal sequence features correspondingly. The obtained features are aggregated and used to predict RUL. According to the experimental findings, the suggested model performs better on subsets FD001, FD002, and FD003 than the state-of-the-art (SOTA) methods. The RMSE evaluation metric shows a reduction of 0.95%, 2.03%, and 1.36%, respectively, while the Scores evaluation metric shows a reduction of 2.53%, 54.89%, and 20.59%. These improvements effectively mitigate the risk of delayed prediction.

Topics & Concepts