Remaining useful life prediction for turbofan engines using an attention-based data-driven deep-learning approach

The remaining useful life (RUL) estimate is the fundamental building block of prognostics and health management (PHM), a field that has the potential to increase system safety and save maintenance expenses. Because of their adaptable structures and improved effectiveness in handling nonlinear behaviors, a variety of deep learning (DL) algorithms have surfaced for RUL forecasting. While DL models such as convolutional neural network (CNN) and Long short-term memory (LSTM) have shown promise, they often struggle to capture complex temporal dependencies and to assign appropriate importance to varying sensor features. The LSTM architecture is based on an attention mechanism integrated with an attention block called Dual Attention Block LSTM (DAB-LSTM). Preprocessing the dataset is the first step in the preparatory phase. The piecewise linear degradation approach is then applied to model RUL labels. The addition of the attention mechanism enhances the input features for time steps that are highly correlated with RUL, which helps to improve the network’s feature extraction performance. The adaptive attention method is then included in an LSTM network to effectively collect and evaluate long-term dependencies, use weighted features, and extract feature representations strategies for paying attention that can improve RUL performance by making use of long-term correlations. The Tree-structured Parzen Estimator (TPE) algorithm is used as a hyperparameter optimization method to find the best values for hyperparameters. In the end, weighted characteristics are combined to create RUL values using a fully connected neural network. We have successfully demonstrated the effectiveness of the DAB-LSTM model on the NASA Commercial Modular Aero-Propulsion System Simulation (C-MAPSS) dataset, and the model has demonstrated improved accuracy compared to other competing approaches.