Edge-Server Collaborative System for Real-Time and In-Depth Damage Detection of Wind Turbine Blades using Acoustic Signals
Abstract
Efficient health monitoring is indispensable for the reliable operation of wind turbines. Damage to wind turbine blades, such as cracks and holes, typically generates whistle-like sounds during rotation. This study proposes a two-stage edge-server collaborative system for detecting blade damage using acoustic signals captured by arrays built from commodity microphones. The first stage employs a lightweight attention-based convolutional neural network to run on edge devices for the real-time binary classification to determine whether anomalous sounds are present. Suspicious time segments are stored for further analysis. The second stage uses a time-frequency sound event detection model that employs a detection transformer with an audio spectrogram transformer backbone to identify the time and frequency ranges of sound events via bounding boxes in the spectrograms. Owing to its high computational demand, this in-depth analysis is performed on a server. To validate the proposed system, acoustic signals were recorded intermittently for more than a year using micro-electromechanical system (MEMS) microphones externally attached to wind turbine towers. The models were trained and evaluated on a manually annotated dataset comprising 4,210 audio clips (15 s each) containing 14,420 sound events. The experimental results demonstrated that the binary classification model achieved an area under the receiver operating characteristic curve (AUC) of 0.920, whereas the sound event detection model attained an average precision at a 50% intersection-over-union threshold (AP50) of 0.510. Furthermore, evaluations on test data under unseen conditions, comprising 496 clips with 135 sound events recorded by handheld recorders at different locations, yielded an AUC of 0.867 and an AP50 of 0.440. The results highlight the robustness of the proposed system to variations in microphone types, recording locations, and environmental noise, demonstrating its strong potential for practical continuous automatic damage detection in wind power infrastructure.