Assessment of gear failure severity in wind turbines based on few-shot learning and graph neural networks

Abstract Distinguishing vibration signals associated with different levels of gear damage, combined with the challenging operating environment of wind turbines, complicates the collection of sufficient data for effective fault diagnosis. This paper proposes a few-shot learning (FSL) based graph neural network (GNN) for evaluating gear tooth fracture levels within small datasets.The short-time fourier transform (STFT) is used to convert the original signal into two-dimensional data for preprocessing. Fault severity features, extracted by a convolutional neural network (CNN), are then input into the GNN for severity classification. A natural wind turbine experiment platform was developed to simulate various operating conditions. To further validate the proposed method, comparative experiments were conducted using Siamese Networks, Matching Networks, and Relation Networks. The results demonstrate that the proposed method outperforms these alternatives in evaluating gear damage severity.