Research on Anomaly Detection of Civil Aircraft Hydraulic System Based on Multivariate Monitoring Data

The hydraulic system is one of the most important systems of an aircraft, and the stainability and reliability of this complex dynamical system directly affect operation and safety. To secure normal operation of the hydraulic system and to assure successful missions, accurate health monitoring, rapid fault diagnosis and isolation, and effective maintenance measures are needed. As a highlight of the new civil aircraft model, some methods based on analysis of a single parameter have been implemented to anomaly detection in Aircraft Condition Monitoring System (ACMS) records for early warning of out-of-gauge events and trend monitoring of performance parameters. Unfortunately, because of the complexity and the coupling relationship between components of the hydraulic system, it is almost impossible to monitor the health state of the hydraulic system via a single variable. To serve the purpose, a comprehensive anomaly detection method considering multivariate monitoring data is proposed in this paper. The unsupervised autoencoder (AE) model with a multi-layer network structure is used for feature extraction of multivariable time series data, and the distance based on similarity is measured to characterize the difference between observations and reconstructions. A health baseline is obtained from a large number of normal training samples and used for anomaly detection and analysis of a new observation. In this paper, a case of driving pump anomaly detection of the hydraulic system is used to illustrate the effectiveness of the proposed method.