Research on Few Shot Anomaly Detection Technique for Wind Turbines Based on Audio Data

Abstract

At present, The predominant method of anomaly detection for wind turbines remains manual "sound inspection" conducted by experienced operation and maintenance personnel, despite its low efficiency and poor timeliness. The application research on wind turbines anomaly detection is still in its early stages, despite the inherent characteristics of security, continuity, and stability associated with acoustic fault detection. Recognizing the audio fault detection characteristics , this paper presents a few shot anomaly detection method for wind turbines based on audio data, to support quick diagnosis by operation and maintenance staff. Firstly, a wind turbine audio fault detection system is established and the collected wind turbine fault audio data is preprocessed using a recurrent neural network (RNNoise). Then, in view of the deficiency of effective samples of wind turbine faults and the presence of pseudo samples, the compression and reconstruction anomaly detection method was used to facilitate few shot anomaly detection. Finally, pseudo faults were further distinguished by the reverse gradient transfer of anomaly data, leading to an enhancement of fault detection accuracy. Actual data collected in an experiment show a more than 95% recognition rate for abnormal sound arising from wind turbine yaw, demonstrating the viability of the proposed technique for deployment at wind farms. This application contributes to improving the intelligent technical level of wind farms.

Keywords

• wind turbines
• anomaly detection
• audio data
• deep learning
• few shot