Remaining Useful Life Estimation Enhancement via Deep Adaptive Feature Extraction

The length between the current point in the degradation process and the time of reaching the failure threshold, or the remaining usable life (RUL) prediction of systems, is of the greatest priority in the industry. More accurate estimation is useful for maintenance decisions as it helps to avoid catastrophic breakdowns and may also assist in reducing additional costs. Deep learning approaches have made impressive advancements in this field in recent years by becoming widely attractive and employed. However, most deep learning approaches don’t fully consider the information implications of sensors adaptively. To overcome this problem, a novel adaptive hybrid model that combines a convolutional neural network (CNN) and gated recurrent unit (GRU) is introduced in this work. The RUL estimation is based on the best practical option of sequence data through CNN-GRU. In the first step, optimal sensor selection is applied to the dataset to collect the most useful sensors. Then, the input data is transformed into a predefined range of values using standard and min-max scalars; in the next step, the normalized data is fed into the CNN-GRU model with an adaptive activation function for deep feature extraction, training, and RUL prediction. Utilizing CNN to extract features from the multivariate input data automatically, the features are then fed into the GRU layer to train the model for RUL prediction. To test the effectiveness of this framework, the suggested methodology is applied to the NASA Commercial Modular    Aero-Propulsion System Simulation (C-MAPSS) dataset. The findings demonstrate that CNN-GRU is capable of accurate RUL prediction. In addition, CNN-GRU outperforms CNN-LSTM and CNN-RNN in terms of computation efficiency and accuracy.