MACHINE LEARNING FOR HEALTH MONITORING AND MANAGEMENT OF POWER PLANTS

Abstract

Power plant plays an important role in modern society. Health monitoring and management is essential to ensure high system availability as well as mitigating the economic losses caused by failure. However, the health monitoring and management for power plant is not trivial due to some challenges arising from several aspects, such as the complex structure of power plant and the diversity of information sources. Motivated by some practical problems, this dissertation aims at developing some machine learning methods of condition monitoring as well as maintenance decision-making for power plants.

We first develop a condition monitoring and fault isolation system for wind turbine based on supervisory control and data acquisition (SCADA) data. This system solves some inherent challenges in SCADA data analysis, including the low sampling rate, time-varying working conditions of the wind turbine, and a lack of historical fault data. We then propose a covariate-regulated sparse subspace learning model to analyze the multivariate data with complex cross-correlation. This model can be applied to the monitoring of multivariate process, which is common in power plant. The dissertation also propose an optimal multi-type inspection policy for gas turbine system. This policy helps to decide on when operators should make an inspection and which kind of inspection should be performed based on the information from online monitoring. These proposed methods show effectiveness in monitoring the health condition of power plant or significant reduction of operational costs, which exhibits the contribution of this dissertation.