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

Various physical characteristics, including ultrasonic waves, active acoustic emissions, vibrations, and thermal imaging, have been used for blade fault detection. In this work, we propose using the sound produced by spinning wind blades to identify faults. To the best of our knowledge, passive acoustic information has not yet been explored for this task. In particular, we develop three networks targeting different scenarios. The main contributions of this work are threefold. First, when normal and aberrant data are available for supervised learning, an attention-convolutional recurrent neural network is designed to show the feasibility of using passive sound information to conduct fault detection. Second, in the absence of abnormal training data, we build a normal-encoder network to learn the distributions of normal data through semisupervised learning, which avoids the requirement of abnormal training data. Third, when multiple devices are used to collect the data, due to different properties of devices, there is a domain mismatch issue. To overcome this, we create an adversarial domain adaptive network to close the gap between the source and target domains. Acoustic signal datasets of actual wind turbine operations are collected to evaluate our fault detection systems. The findings demonstrate that the proposed systems offer high classification accuracy and indicate the feasibility of passive acoustic signal-based wind turbine blade fault detection with one step close to automatic detection.