Label-smoothing dynamic decoupling augmented network for intelligent fault diagnosis under imbalanced data distribution with noisy labels

Engineering applications frequently face imbalanced data distributions due to rare fault types, resulting in a higher prevalence of data from dominant fault classes. This imbalance causes deep learning models to be biased toward majority classes while inadequately representing minority classes. Noisy labels, stemming from data collection errors, annotation mistakes, and data handling issues, introduce incorrect supervision in deep learning networks, consequently affecting fault diagnosis performance. To address above challenges, a dynamic decoupling augmentation network with label aware smoothing (LS-DDAN) is proposed for imbalanced data fault diagnosis under noisy labels. Specifically, the feature extractor and the classifier are dynamically decoupled, enhancing feature representation and optimizing classifier training under imbalanced data distribution. Then, a label-aware smoothing augmentation framework guided by noisy labels perception is introduced to mitigate the influence of noisy labels. This framework adaptively smooths and reconstructs noisy labels, reducing their confidence while directing the model's attention toward critical fault features. Finally, by calculating the effective volumes of different classes, the classification boundaries are further refined. Experimental results across two datasets show that the proposed method achieves an average improvement of approximately 2% in recognition accuracy compared to state-of-the-art methods for fault diagnosis under imbalanced data distributions with noisy labels.