Unsupervised Fault Detection Driven by Multivariate Time Series for Aeroengines

Abstract

Aeroengine fault detection is extremely significant to ensure flight safety. However, it still poses a vast challenge due to the complexity of aeroengine systems. In this paper, an unsupervised deep autoencoder (DAE) with dimension fusion function (DFF) method (DFF-DAE) was proposed to effectively detect aeroengine faults based on multivariate time-series data. The proposed method does not rely on mathematical models and expert experience to automatically extract the hierarchical representation features of raw data. The DFF-DAE has a three-layer autoencoder with its first encoding layer incorporating a DFF, which is developed for multivariate time-series fusion before encoding. A case study of aeroengine rolling bearing locking fault was conducted on the turbofan engine ground test data composed of vibration and rotational speed time-series data in variable working states. The results indicated that DFF-DAE succeeds in efficiently employing multivariate time-series data to perform accurate fault detection. Compared with other state-of-the-art methods, DFF-DAE yields superior predictive ability and detection accuracy. Furthermore, DFF-DAE provides application potential for online fault detection because it is effective in variable working states of the aeroengine.