An Anomaly Detection Approach Based on Autoencoders for Condition Monitoring of Wind Turbines

This research presents an anomaly detection approach based on autoencoders for wind turbine condition monitoring. The overall goal is to develop a methodology for assessing wind turbine health in order to enable preventative maintenance programs. First, SCADA signals are used as data input in the approach. The approach then examines the differences between the estimated values by the autoencoder models and the measured signals from the SCADA system. Next, the Kernel Density Estimation is used to determine the distribution of the expected output’s error. Finally, to efficiently extract anomalous activity in the data, a novel dynamic thresholding approach is used. The outcome reveals that the method is capable of detecting potential abnormalities prior to the onset of a breakdown. It also verifies that the approach can alert operators to potential changes within wind turbines even when the alarm records show no alerts.

I. Introduction

In the last decade, technological advances have resulted in an increase in the use of renewable energy (RE). According to International Energy Agency (IEA) reports, demand for fossil fuels has declined, while the use of RE has increased by about 3%. While COVID-19 problems have resulted in financial losses for many organizations, the degree of dedication to RE continues to be astounding. The IEA predicted that renewable energy generation will expand by over 8% in 2021, reaching 8,300 TWh, the highest annual rise since the 1970s [1], [2]. Further, solar and wind energy are expected to account for two-thirds of renewable growth. Two years after the pandemic began, the wind sector has shown amazing resilience in the face of unprecedented challenges. Despite employment losses and project delays, the global wind business has defied expectations and is set to grow at a healthy rate. According to GWEC Market Intelligence, the sector’s total onshore and offshore wind installations will exceed 1TW by 2025 [2], [3].