Remaining useful life prediction for bearing of an air turbine starter using a novel end-to-end network

Abstract An air turbine starter (ATS) is used to start an aircraft’s engine before the aircraft takes off, as part of which the rolling bearings are an essential but easily damaged component. Predicting the remaining useful life (RUL) of an ATS bearing is a key part of efficient engine maintenance. To improve the prediction accuracy of rolling bearings’ working in complex environments, this paper proposes a novel end-to-end network for RUL prediction based on parallel convolution and a bidirectional long and short-term memory (BiLSTM) network. The architecture is an integration of two parts: feature extraction and RUL prediction. For the feature extraction, a more tailored one-dimensional convolution neural network architecture has been adapted for multi-rate sensors in a parallel manner, and a multiscale feature stacking and mixing mechanism is further designed following the convolution operation to extract the most representative degradation feature. In the prediction part, environmental factors are added to the BiLSTM network together with the previously extracted degradation feature. Both parts of the end-to-end network can focus on valuable information without any prior knowledge due to utilization of an attention mechanism. A real data set is built to evaluate the performance of the proposed method, and the RUL predictive error percentage decreases by 1.02% compared with the existing algorithm.