Multimodal Data Enhanced Wind Generator Fault Diagnosis Considering Small Sample Conditions

Data-driven methodologies are effective tools for diagnosing frequent occurrences of static air-gap eccentricity (SAGE) and stator interturn short circuit (SISC) faults of generators. Nonetheless, these approaches exhibit pronounced dependence on the availability and volume of fault samples, thereby limiting the application of diverse diagnostic models. To tackle these challenges, an innovative fault diagnosis model for wind generators based on Wasserstein deep convolution generative adversarial network with gradient penalty (WDCGAN-GP) and double-branch wavelet convolutional neural network (DWCNN) is proposed. Initially, the phase current and stator vibration are selected as the diagnostic media based on the fault mechanism analysis. In addition, the WDCGAN-GP incorporates Wasserstein distance (WD) and gradient penalty (GP) to enhance training stability and generate high-fidelity current and vibration data to augment the limited sample dataset. Subsequently, the DWCNN is adopted to undertake the diagnosis task, which is achieved through training the model using the extended dataset. Finally, the proposed methodology is experimentally validated on a 13 kW external rotor permanent magnet generator (ERPMG). The results substantiate that the proposed method achieves precise fault diagnosis based on generating high-fidelity time-domain signals under limited sample conditions.