Remaining Useful Life Prediction via a Double Convolutional Attention-Based CNN-GRU Model

Accurate prediction of the remaining useful life (RUL) of the important equipment in the industrial is crucial for making operational planning and maintenance decisions. Recently, RUL prediction models such as convolutional neural networks (CNNs) and gated recurrent neural networks (GRUs) have demonstrated excellent RUL prediction performance. However, CNNs and GRUs are unable to inherently implement adaptive weighting of multi-sensor data. Moreover, the integrated methods with CNNs and GRUs have shown obvious limitations in terms of prediction accuracy. To overcome these problems, a double convolutional attention-based CNN-GRU model (DCAB-CNN-GRU) is proposed in this paper. Firstly, the sensor signals are preprocessed and the RUL labels are strategically modeled using the piecewise linear degradation principle. Subsequently, a double convolutional attention-based CNN-GRU model is utilized to predict RUL. As a core of our approach, the double convolutional attention mechanism which is devoid of complex structure and computational complexity, can dynamically assign weights to features based on the salient time points within the degradation trajectory, thereby enhancing sensitivity of the model to the critical degradation information. Finally, the experimental results using aircraft engine datasets verify the effectiveness of the proposed DCAB-CNN-GRU over the state-of-the-art methods.