Adaptive Weighted Cost-Sensitive Learning-Driven Improved Dense Convolutional Neural Network for Imbalanced Fault Diagnosis under Limited Fault Samples

Abstract

Offshore wind turbines play a crucial part in the transformation of wind energy into electricity, which significantly benefits the sustainable development of the economy and society. Nevertheless, offshore wind turbines in practice are often in extremely severe operating environments, giving them tremendous challenges for their safe operation. In particular, the scarcity of fault data in the actual operating scenarios makes it difficult to collect enough fault data for training, resulting in a long-tailed distribution of training data, which leads to the majority class dominance and minority class overfitting problems. For the above-mentioned problems, an adaptive weighted cost-sensitive learning-driven improved dense convolutional neural network is proposed. First, a large convolutional kernel and interactively replicated dense connections are utilized to extract more stable discriminative features with fewer parameters. Second, an activation function with self-normalization property enhances the stability of model training under imbalanced data conditions. Further, adaptive weighting of misclassification cost is achieved by integrating sample size distribution, sample importance information, and imbalanced classification assessment metrics. Finally, two cases and ablation experiments under the wind turbine simulator testbed are implemented to validate the effectiveness and superiority of the proposed method.