Multivariate Anomaly Detection and Early Warning Framework for Wind Turbine Condition Monitoring Using SCADA Data

Abstract

Wind speed power characteristics are essential in evaluating the state of the wind turbine. The supervisory control and data acquisition (SCADA) data are massively collected and could be important resources for condition monitoring and anomaly detection of wind turbines if properly utilized. A systematic early-stage anomaly detection framework is built in this work consisting of three phases: (1) an improved data cleaning algorithm based on kernel density estimation (KDE) is presented to remove outliers of SCADA data where the constraint of the Gaussian distribution assumption is eliminated for describing the real distribution of power outputs in each wind speed interval; (2) deep neural networks (DNNs) are used to establish a multivariate power curve (MPC) model where the dependencies of multidimensional variables on power output are considered and selected by Pearson correlation analysis; and (3) the sequential probability ratio test (SPRT) is adopted to estimate the distribution of power residuals and used for anomaly detection and early warning. The case studies verified the efficacy of the proposed framework where 91 faults from 38 wind turbines in two wind farms are successfully detected in the early stage.