A Data Analytics Approach to Failure Precursor Detection of Gas Turbine

Abstract

In this paper, a novel approach is proposed to detect precursory events that lead to catastrophic systems failures. This approach is applied to investigating failures of heavy duty gas turbines. Current industry standards rely on either vibration sensors or gas path performance measurement sensors to identify system anomalies, but this proposed process is based on a combination of information from both type of monitoring sensors. This process is built on a systematical multi-step concept developed by assembling proven mathematical and statistical signal processing techniques to achieve a robust and more precise failure precursor detection methodology. The first step includes performing a multi-resolution analysis of gas turbines gas path performance measurement parameters, condition monitoring and vibration sensors data using wavelet packet transform to extract their signal features. Then, the probabilistic principal component analysis is utilized to fuse data of different types into a set of uncorrelated principal components. Next, a one-dimensional signal representing the multi-variable data is computed. After that a statistical process control technique is applied to set the anomaly threshold. Finally, a Bayesian hypothesis testing method is applied to the monitored signal for abnormality detection. As a proof of concept, the proposed process is successfully applied to a gas turbine compressor failure precursor detection problem.