A novel wind turbine bearing fault diagnosis method based on High-Order Spectrum Analysis (HOSA) and YOLOv12

Abstract Due to the growing popularity of wind energy, its construction reliability is becoming a high priority engineering issue regarding the maintenance of wind turbine parts, mainly, bearings. This paper proposes a combined fault-diagnosis pipeline that can integrate High-Order Spectral Analysis (HOSA) object detection framework YOLOv12. The bearing data set based on vibration data was transformed into a set of bispectrum images through HOSA, which enhanced the phase coupled non-linear components. We used these bispectrum images to train YOLOv12 on a multi-class task of bearing conditions including Normal, Inner Race Fault, Outer Race Fault and Ball Fault. The experimental setup is very simple, a model is trained on 2000 bispectrum images (500 images per class), and standard measurements of detection performance are applied. The proposed system produced a 99.90% precision and a 100% recall with a 99.5% of mean Average Precision (mAP@0.5). Such outcomes demonstrate the almost perfect classification accuracy. The independency and generalization ability of the model on all the fault types was shown further through F1-confidence analysis, precision–recall curves, and confusion matrices. The HOSA-YOLOv12 integration is used to achieve an effective, fast, and scalable solution of detecting faults in bearings of wind turbines accurately. This research proves that deep-learning powered spectral diagnostic is feasible and establishes a model of intelligent and condition-based maintenance of renewable energy systems.