A Deep Parallel Spatiotemporal Network Based on Feature Cross Fusion for Remaining Useful Life Prediction of Aero Engine

Accurately predicting the remaining useful life (RUL) of aero engines can make maintenance plans in advance and effectively avoid serious aviation accidents caused by engine failures. In the face of multi-sensor data, the existing deep learning often uses a single feature of time or space for RUL prediction, which fails to fully consider the parallel extraction of time and space dimension features and the effective fusion of features of various dimensions. To this end, this paper proposes a deep parallel spatiotemporal network based on feature cross-fusion (FCF-DPSTN) to effectively fuse spatiotemporal features and improve the accuracy of RUL prediction. Firstly, the parallel structure composed of enhanced gated recurrent unit (EGRU) and FNet is used to mine long-term dependency features and global frequency features on the time axis, while the proposed dynamic convolutional (Dy conv) layer is used to mine local spatial features on the spatial axis. Subsequently, the feature cross-fusion module integrates multi-angle information to dynamically weight the contribution of features of different dimensions. Finally, the feature regression module uses the approximate Bayesian method to realize the RUL interval estimation of the engine. The proposed method is validated on two aero-engine datasets, and the experimental results show that the prediction performance of the proposed method is better than that of the advanced prediction methods.