RGNet: Renormalization Group Neural Network for Imbalanced Fault Diagnosis.

Machine learning models often struggle with practical challenges such as class imbalance and high-dimensional noise, which can significantly degrade performance, particularly in critical applications like fault diagnosis. To address these issues, a new neural network architecture called RGNet has been proposed. RGNet is built upon the principles of the renormalization group (RG), enabling a hierarchical coarse-graining of the feature space. This design allows the model to sequentially compress input dimensionality while retaining information from various scales, ultimately concatenating these multi-scale representations for classification. This unique approach helps RGNet effectively capture both fine-grained local details and broader global patterns within the data. The research also introduces "RG-flows," which are interpretable low-dimensional representations derived from the model. Visualizations of these flows confirm the efficacy of the coarse-graining process. Empirical evaluations on the imbalanced AI4I dataset demonstrate that RGNet provides a universal, interpretable, and competitive solution for fault prediction in scenarios characterized by imbalanced classes.