A Novel Frequency Domain Attention Enhanced CNN-Transformer Network for Remaining Useful Life Prediction

Accurate prediction of the Remaining Useful Life (RUL) is indispensable for maintaining system safety, serving as a cornerstone in the domain of Prognostics and Health Management (PHM). Approaches based on neural networks with manually extracted features are overly dependent on specialized knowledge and underutilize monitoring data, often resulting in low estimation accuracy. This study proposes a frequency domain attention-enhanced CNN-Transformer model for RUL prediction. The proposed method begins with automatic health feature extraction by applying a convolutional neural network layer to the time-series data collected by the sensors, which can reduce the dimensionality of the input health feature vector. Then, the Fourier layer allows feeding frequency domain information to multi-head attention. Finally, the Transformer architecture is utilized to reduce computational costs and improve the accuracy of long-term predictions. In this paper, a turbofan engine is selected as a typical engineering system and the effectiveness and performance of the proposed method are validated using NASA's C-MAPSS dataset. The RMSE of RUL prediction results exhibits better performance compare to traditional methods.