A Statistical Wavelet-Based Process for Systems Catastrophic Failure Precursor Detection

Abstract

As a consequence of the recent deregulation in the electrical power production industry, new private entrepreneurs with no prior experience in the power plant operation have entered into the power generation business. They hedge the business risks by outsourcing the operation and maintenance activities to third party service providers with whom they share risks/rewards of plant performance. The main maintenance providers are the original equipment manufacturers, who are responsible for the majority of the cost associated with unplanned outages. With the cost-benefit of preventing such unplanned outage as a gas turbine compressor failure hovering around the twenty million dollars mark, techniques for detecting failure precursors to avoid or limit the number of systems catastrophic failure are necessary. In this paper, a methodical process is proposed to detect precursory events that lead to catastrophic systems failure. The wavelet packet transform is used to perform multi-resolution analysis of gas turbines health, condition and vibration sensors data to extract their signal features. Then the probabilistic principal component analysis is utilized to fuse them into a few uncorrelated variables. Next a one-dimensional signal representing the multi-variables data is computed. After that the statistical process control techniques is applied to set the anomaly threshold. Finally, a Bayesian hypothesis testing method is applied for abnormality detection to the monitored signal. As a proof of concept, the proposed process is successfully applied to a gas turbine compressor failure precursor detection.