Deep Learning for Acoustic Pattern Recognition in Wind Turbines Aerial Inspections

Abstract

Wind turbine maintenance management requires new condition monitoring systems and robust algorithms for fault detection. Acoustic inspection can detect anomalies in rotatory components through the analysis of acoustic data with advanced pattern recognition algorithms. Unmanned Aerial Vehicles can carry large sensors and technologies, being one of the most relevant techniques for non-destructive inspections. This article presents a new approach to analyze the acoustic data acquired by the condition monitoring system developed in previous research studies, formed by an acoustic sensor embedded in an unmanned aerial vehicle. This approach develops different deep learning methods for acoustic signal analysis to detect abnormal patterns. One of the most suitable tools for this purpose is the Recurrent Neural Network, concretely the Long Short-Term Memory neuronal network commonly employed in sound pattern recognition. It is presented a real case study, obtaining a sound pattern classification based on the trained networks achieving 87% accuracy. This methodology aims to develop a future integrated system capable of detecting wind farms anomalies with acoustic datasets.