A novel unsupervised anomaly detection for gas turbine using Isolation Forest

Monitoring gas turbines’ health, in particular, detecting abnormal behaviors in time, is critical in ensuring gas turbine operating safety and in preventing costly unplanned maintenance. One most popular anomaly detection method is to obtain a classification-prediction model by training a classifier using the real-life data of gas turbine. The excellent detection ability of this method is attributed to enough annotated samples, especially enough annotated abnormal samples. Nevertheless, in gas turbine monitoring data, normal data is far more than abnormal data, even no abnormal data. Advanced technologies that can accurately detect the abnormal behaviors in time using the unlabeled data are in great need. Thus, a novel unsupervised anomaly detection based on Isolation Forest is investigated for gas turbine gas path anomaly detection in this paper. Specifically, the monitoring data is grouped by time series for weakening the affection of inevitable performance degradation when gas turbine operating, and then all detected by an isolation forest model with low contamination. Each detected abnormal group is detected again by an isolation forest model with high contamination for obtaining the specific abnormal flight-cycles. Using the real-life monitoring data from 8 different CFM56-7B aeroengines, the detection results show that the method based on Isolation Forest can achieve high accuracy abnormal detection under unlabeled data and small data set.