Multi-Scale Temporal-Spatial Feature-Based Hybrid Deep Neural Network for Remaining Useful Life Prediction of Aero-Engine

Abstract

Remaining useful life (RUL) prediction is crucial for simplifying maintenance procedures and extending the lifespan of aero-engines. Therefore, research on RUL prediction methods for aero-engines is increasingly gaining attention. In particular, some existing deep neural networks based on multiscale features extraction have achieved certain results in RUL predictions for aero-engines. However, these models often overlook two critical factors that affect RUL prediction performance: (i) different time series data points have varying importance for RUL prediction, and (ii) the connections and similarities between different sensor data in both directions. This paper aims to extract valuable multiscale features from raw monitoring data containing multiple sensor measurements, considering the aforementioned factors, and leverage these features to enhance RUL prediction results. To this end, we propose a novel deep neural network based on multiscale features extraction, named Multi-Scale Temporal-Spatial feature-based hybrid Deep neural Network (MSTSDN). We conduct experiments using two aero-engine data sets, namely C-MAPSS and N-CMAPSS, to evaluate RUL prediction performance of MSTSDN. Experimental results on C-MAPSS data set demonstrate that MSTSDN achieves more accurate and timely RUL predictions compared to 12 existing deep neural networks specifically designed for predicting RUL of aero-engine, especially under multiple operational conditions and fault modes. And experimental results on N-CMAPSS data set eventually indicate that MSTSDN can effectively track and fit with the actual RUL during the engine degradation phase.