A Physical-Feature Interactive Expansion-Based Fault Diagnosis Method With Applications to Marine Current Turbines

The problem of imbalanced samples is prevalent in condition monitoring data of marine current turbines (MCT). It brings critical challenges for training data-driven fault diagnosis models and could significantly deteriorate the model performance. To address this issue, a fault diagnosis approach based on the physical-feature interactive expansion (PFIE) and convolutional neural networks (CNN) is developed in this article. The proposed method combines the data augmentation of PFIE with the classification capacity of CNN. Specifically, a PIPE method is first designed to augment real-world samples from MCT physical fault frequencies. The samples are then synthesized via a physical-feature interactive optimization function. The synthetic samples are finally trained by CNN, and online monitoring data are recognized with trained models. Experimental results show that the proposed fault diagnosis approach can achieve high accuracy compared with other approaches.