A Novel Ensemble-Learning-Based Convolution Neural Network for Handling Imbalanced Data

Abstract

Deep-learning-based fault diagnosis of wind turbine has played a significant role in advancing the renewable energy industry. However, the imbalanced data sampled by the supervisory control and data acquisition systems has led to low diagnosis accuracy. Additionally, deep neural networks can encounter issues like gradient vanishing and insufficient feature learning during backpropagation when the model is too deep. This article introduces a novel approach that is based on dynamic weight loss functions to modulate unbalanced data and improve diagnostic accuracy by focusing on misclassification of a small sample number. The proposed approach employs a 1D-CNN model and an ensemble-learning-based convolution neural network (EL-CNN) to enhance diversity of models and complementarity of feature learning. The EL-CNN model addresses the problem of local features being overlooked and provides more accurate results. The effectiveness of this proposed approach is well demonstrated through experimental cases on real wind turbine pitch system fault data. Two different networks using three different loss functions and three state-of-the-art fault diagnosis models are tested, demonstrating the EL-CNN model’s superiority.