A Weakly Supervised Learning-Based Oversampling Framework for Class-Imbalanced Fault Diagnosis

With the lack of failure data, class imbalance has become a common challenge in fault diagnosis industrial systems. The oversampling methods can tackle class-imbalanced problem by generating minority samples to balance training set. However, one main challenges existing is how generate high-quality samples. Traditional regard all synthetic as ones be added set without filtering. low-quality would distort distribution dataset and worsen classification performance. In this article, we propose weakly supervised method that treats unlabeled develops graph semisupervised learning algorithm select samples, adding into final To improve quality cost-sensitive neighborhood component analysis dimensionality reduction enhance domain information validity high-dimensional datasets. Finally, combining boosting-based ensemble framework, new imbalanced framework suitable for high highly experimental validation performed on five real-world wind turbine blade cracking datasets compared 15 benchmark methods. results show average performances robustness proposed are significantly better than those