Nonintrusive wind blade fault detection using a deep learning approach by exploring acoustic information

Area of Science:

• Engineering
• Acoustics
• Artificial Intelligence

Background:

• Traditional wind turbine blade fault detection methods rely on physical characteristics like vibrations or thermal imaging.
• Passive acoustic monitoring, analyzing the sound emitted by spinning blades, has not been previously explored for fault detection.

Purpose of the Study:

• To develop and evaluate novel machine learning models for wind turbine blade fault detection using passive acoustic signals.
• To address challenges including supervised learning with limited abnormal data and domain mismatch across different data acquisition devices.

Main Methods:

• An attention-convolutional recurrent neural network for supervised fault detection when normal and abnormal data are available.
• A normal-encoder network utilizing semi-supervised learning to detect faults without abnormal training data.
• An adversarial domain adaptive network to mitigate domain mismatch issues when using data from multiple devices.

Main Results:

• The proposed methods achieved high classification accuracy in detecting wind turbine blade faults using passive acoustic data.
• Demonstrated the feasibility of using sound produced by spinning blades for effective fault identification.
• Successfully addressed scenarios with limited abnormal data and domain variations.

Conclusions:

• Passive acoustic signal analysis is a viable and effective approach for wind turbine blade fault detection.
• The developed deep learning models offer robust solutions for different data availability and acquisition scenarios.
• This research represents a significant step towards automated wind turbine condition monitoring.