Neural Network Implementation of the Renormalization Group for Fault Diagnosis with Class Imbalance
Researchers have developed RGNet, a novel neural network architecture inspired by the renormalization group (RG) concept, designed to address challenges like class imbalance and multidimensional noise in machine learning applications. This model achieves hierarchical coarse-graining by sequentially compressing feature space dimensionality and concatenating features across different scales before classification. The RGNet's effectiveness is demonstrated through interpretable low-dimensional representations visualized with t-SNE, revealing a discrete curvilinear structure that confirms its coarse-graining capabilities. Applied to the imbalanced AI4I dataset, RGNet proved to be a versatile, interpretable, and competitive solution for fault prediction tasks. AI
IMPACT Introduces a new architecture for handling class imbalance and noise in fault diagnosis, potentially improving model interpretability and performance.