Feature selection and adaptive threshold for automated cavitation detection in hydroturbines

Abstract

Hydroturbines produce 6.3% of all electrical generation and 48% of renewable energy in the United States of America. While hydro power plants have existed for well over 100 years, cavitation damage on hydroturbine runners remains as an expensive problem that reduces power production and shortens the life of the turbine. Hydroturbine operators who wish to perform cavitation detection and collect intensity data for estimating the remaining useful life (RUL) of the turbine runner face several practical challenges related to long term cavitation detection. This thesis presents both a method for comparing and evaluating cavitation detection features as well as a method for creating adaptive cavitation thresholds and automating the cavitation detection process. The method for cavitation feature selection can be used to quickly compare features created from cavitation survey data collected on any type of hydroturbine, sensor type, sensor location, and cavitation sensitivity parameter (CSP). Although the cavitation feature selection process is based on manual evaluation and knowledge of hydroturbine cavitation, the use of principal component analysis greatly reduces the number of plots that require evaluation. A case study based on data taken from a production hydroturbine is used to demonstrate the method and the results provide a clear ranking of the preferred sensors, sensor placements, and CSPs for the hydroturbine - thus demonstrating the usefulness of the method. The the second method presented in this thesis addresses several challenges encountered when detecting cavitation for long periods of time { a prerequisite to developing a datadriven method for estimating cavitation erosion rates. First, adaptive cavitation thresholds are generated by collecting sensor data from a hydroturbine ramp-down, then creating CSPs from the data and calculating the Mahalanobis distance (MD) to create clear separation between the healthy running state and conditions where the hydroturbine is experiencing cavitation. Next, in order to automate the cavitation detection process, the cavitation threshold is used to create class labels for the ramp-down data which is then used to train a supervised learning algorithm for classifying cavitation from sensor data. Although domain knowledge is still required to select appropriate CSPs, the remainder of the process can be automated by applying unsupervised learning to label the training set. This method is also demonstrated utilizing data collected on production hydroturbines in a power plant environment. The results of the case studies indicate that the fully automated process for selecting cavitation thresholds and classifying cavitation performed well when compared to manually selected thresholds. Our methods provide hydroturbine operators and researchers with a clear and eective way to perform automated cavitation detection while also laying the groundwork for determining RUL in the future.