Long-Tailed Multi-Label Diagnosis of Compound Faults in Wind Turbine Gearboxes via Multi-Channel Imaging of FBG Vibration Signals

Abstract

Wind power plays an important role in renewable energy generation, and the reliability of wind turbine gearboxes directly affects turbine operation and maintenance. Compound gear fault diagnosis remains challenging because multiple fault components may coexist and compound fault samples are often limited, leading to long-tailed data distributions. To address this problem, this study proposes a long-tailed multi-label diagnostic framework based on fiber Bragg grating (FBG) acceleration signals and multi-channel time-series imaging. Missing tooth, pitting, and tooth breakage faults are encoded as three independent labels to represent healthy, single-fault, double compound fault, and triple compound fault conditions. The one-dimensional FBG wavelength-shift signals are transformed into GASF-GADF-MTF three-channel images, which describe amplitude angular correlation, dynamic angular difference, and state transition information. A ResNet18-SE network trained with Focal Loss is developed to improve the recognition of minority compound fault samples. Experimental results show that the proposed method achieves an Exact Match Accuracy of 0.9950 and a Macro-F1 of 0.9980 on the Balanced dataset. Under the severe LT50 setting, it achieves an Exact Match Accuracy of 0.9739 and an F1123 of 0.9469. These results demonstrate the effectiveness of the proposed framework for FBG-based long-tailed compound fault diagnosis.