PITA FY2014 Projects

Probabilistic condition assessment for wind turbine components

PI:Matteo Pozzi, Dept. of Civil and Environmental Engineering

University:Carnegie Mellon University

Co-PI(s):J. Zico Kolter, School of Computer Science

Industry Affiliate(s):EverPower Wind Holding, Inc.

Summary:This project will develop an approach to process the monitoring data of a wind farm, at system level, for detecting anomalies and predicting the residual life of components in the turbines. The approach will be based on probabilistic analysis using Dynamic Bayesian Networks (DBN), and make use of machine learning algorithms for inference and learning.

The research will be conducted in collaboration with EverPower Holding Inc., which owns and operates many farms inside and outside Pennsylvania. Their wind turbines are instrumented with several sensors and their database has a high potential for research and educational purposes, as it contains data sets related to healthy and to damaged components, it refers to a long time span, it includes a wide range of environmental conditions and data are collected simultaneously from several similar components.

We will make use of the DBN framework to model the time dependency among the state of each component, features extracted from vibration data and environmental data such as temperature and wind condition. After having learnt the model from the vast database, we will be able to predict the future evolution of the component state, and update the prediction depending on the available measurements. This process will be conducted at the level of the entire farm, combining data collected on similar components.

The first outcome of the project will be a database of classified component responses, depending on the condition state and the environmental condition. The main outcome, however, will be an analytical and numerical model for probabilistic data processing for a family of turbine components.

The project will contribute to the education of students at CMU, by providing data and models useful for many courses offered by the PI and Co-PI.