Detection of Long Term Vibration Deviations in GasTurbine Monitoring Data

Abstract: Condition based monitoring is today essential for any machine manufacturer tobe able to detect and predict faults in their machine fleet. This reduces the maintenancecost and also reduces machine downtime. In this master’s thesis twoapproaches are evaluated to detect long term vibration deviations also called vibrationanomalies in Siemens gas turbines of type SGT-800. The first is a simplerule-based approach where a series of CUSUM test are applied to several signalsin order to check if the an vibration anomaly has occurred. The secondapproach uses three common machine learning anomaly detection algorithm todetects these vibration anomalies. The machine learning algorithms evaluatedare k-means clustering , Isolation Forest and One-class SVM. This master’s thesisconclude that these vibration anomalies can be detected with these ML modelsbut also with the rule-based model with different levels of success. A set of featureswas also obtained that was the most important for detection of vibrationanomalies. This thesis also presents which of these models are the best suitedanomaly detection and would be the most appropriate for Siemens to implement.