Data-Driven Anomaly Detection Methods For Aircraft Engine

Aero engine is a complex equipment which may operates under several extreme conditions.The stability of aero engine has a significant impact on the safety and reliability of the aircraft.The detection and analysis of anomaly conditions in aero-engine is an important method to ensure the safety.In order to ensure the timely detection of anomalies which have occurred or may occur in aero engine during the maintenance or operation,it is necessary to find more efficient and accurate anomaly detection methods.The research contents in this paper includes: The anomaly detection in the value of aero engine key parameters,the anomaly detection in aero engine time series data and the pre-warning of aero engine anomaly state based on the degraded mode.This three research topics cover different aspects of aero engine anomaly detection.The main research contents and innovations of this paper are as follows:(1)The anomaly detection in the value of aero engine key parameters focuses on detecting the anomalies in the value of each key parameter of the aero engine.In order to solve the problem that the dimension of aero engine data is too high,the anomaly feature selection algorithm AFS-IC based on mutual information and cooperative game is proposed.The algorithm finds the best feature subset by calculating the correlation degree among each attribute and anomaly.At the same time,because of the unbalancing of aero engine data,the weighted isolation forest algorithm WIForest is proposed based on the original isolation forest algorithm,which gives weight to different anomaly features and improves the effect of anomaly detection algorithm under high dimensionality and data imbalance.Experiments show that the combination of AFS-IC algorithm and WIForest algorithm has higher precision in the anomaly detection of aero engine’s key parameter value.(2)The anomaly detection in aero engine time series data focuses on the detection of anomaly sequences in aero engine timing data.In order to solve the problem that the time series data contains few abnormal samples,an unsupervised anomaly detection algorithm TFDTW-JDD is proposed.The algorithm uses the Jth-nearest neighbor anomaly algorithm to solve the "Freak Twins" problem caused by continuous abnormal data in time series data.At the same time,the dynamic time warping algorithm with local trend feature is used to optimize the similarity calculation of time series data in the Jth-nearest neighbor anomaly algorithm,which achieved higher accuracy of anomaly sequence detection.(3)The pre-warning of aero engine anomaly state based on the degraded mode focuses on measuring the state and health of the aero engine,and by fitting its state degradation,it can give us pre-warning of possible aero engine anomaly conditions.In order to solve the problem that the degenerate trajectory trend is not considered in the degenerate trajectory similarity based TSBAP algorithm,an anomaly pre-warning algorithm IF-TSBAP based on multi-information fusion similarity measure is proposed,and the original TSBAP algorithm is optimized from the perspective of similarity measure.Experiments show that the proposed algorithm can get higher precision comparing with the original TSBAP algorithm,and IF-TSBAP is better than the algorithms based on other prediction models,which may prove that IF-TSBAP has high accuracy and practicability.