Using audio-data for anomaly detection in the fatigue test of a composite tidal turbine blade

Abstract

FastBlade is a research facility for testing large-scale composite and metal structures. Fatigue tests run on tidal turbine blades measure the mechanical response of a blade subject to the number of loading cycles that mimic the ones it will experience over its lifetime of a subsea deployment. To maximise its throughput by running the facility uninterruptedly, unmanned operation of the site should be possible. One of its key enablers is anomaly detection. Microphones are used as a non-specific and affordable sensing method. Using the audio data, we applied a Fast Continuous Wavelet Transform to extract the patterns recorded during normal and abnormal operations. These outputs are used to train a neural network autoencoder (NNA). The original image is reconstructed from the compressed vector in the latent space (LS) of the NNA, and the loss is computed to detect and quantify anomalies. The study's findings demonstrate the success of using audio data to detect short-lived anomalies despite limited information about the critical assets in the set-up and can be easily extrapolated to other systems.