Acoustic Camera-Based Anomaly Detection for Wind Turbines

Health monitoring of wind turbines (WTs) has gained a lot of attention recently. Prevalent solutions mainly rely on the status data from Supervisory Control And Data Acquisition (SCADA), which is widely installed on modern WTs, to detect failures. However, the sensor data from SCADA may not be sufficient to identify defects of blades. One possible approach is to collect audio signals from the target WTs. Nonetheless, lacking spatial information, audio signals are unable to pinpoint the locations of anomalies. In this work, we propose to employ acoustic imaging for WTs anomaly detection. A reconstruction-based anomaly detection model with a Spatial-Temporal Convolutional Autoencoder is developed. The core idea is to learn the visual representations of acoustic images from a healthy state and therefore a poor reconstruction result would indicate an anomaly. To the best of our knowledge, this is the first attempt to adopt acoustic imaging to handle anomaly detection in the field of WTs. Experimental results demonstrate the effectiveness and robustness of the proposed method across various anomalous conditions.