RGNet: Neural Network Architecture Using Renormalization Group for Imbalanced Fault Diagnosis

Researchers introduced RGNet, a neural network architecture grounded in renormalization group theory, designed to handle class imbalance and multidimensional noise in fault diagnosis tasks. The model uses hierarchical coarse-graining and interpretable RG-flows to capture patterns at multiple scales, demonstrating competitive performance on the imbalanced AI4I dataset.

Quick Facts

Who

Machine learning researchers

What

Proposed RGNet neural network architecture

When

Submitted on 16 June 2026

Where

arXiv Computer Science repository

Proposed RGNet neural network architecture
Applied renormalization group concept to machine learning
Implemented hierarchical coarse-graining of feature spaces
Introduced RG-flows for interpretable representations
Conducted experimental validation on AI4I dataset

Researchers have proposed RGNet, a novel neural network architecture that applies the renormalization group concept from physics to address class imbalance and noise challenges in machine learning fault diagnosis tasks. The architecture implements hierarchical coarse-graining of feature spaces, sequentially compressing input dimensionality while concatenating representations at multiple scales to capture both local details and global patterns simultaneously.

The key innovation of RGNet involves the introduction of RG-flows, which are interpretable low-dimensional representations of the data. Visualization of these representations using t-SNE analysis reveals a discrete curvilinear structure that confirms the effectiveness of the coarse-graining process. This interpretability feature distinguishes RGNet from many black-box machine learning approaches, providing users with insights into how the model processes and transforms data at different levels of abstraction.

Experimental validation was conducted on the imbalanced AI4I dataset, a benchmark used for evaluating fault prediction models in datasets with unequal class distributions. The results demonstrate that RGNet provides a universal, interpretable, and competitive solution for fault prediction in applications where class imbalance is a significant challenge. The approach addresses a persistent problem in practical machine learning applications, where real-world data often features skewed distributions that can degrade model performance and lead to biased predictions.

Topics

Technology Tech Breakthrough Science Artificial Intelligence

#coarse-graining #neural networks #renormalization group #machine learning #feature compression #interpretable machine learning #class imbalance #fault diagnosis

Why This Matters

RGNet addresses a critical challenge in real-world fault diagnosis: class imbalance. In industrial and engineering applications, failure cases are often significantly rarer than normal operation, causing standard machine learning models to underperform. By combining physics-inspired renormalization group theory with neural networks, RGNet provides both competitive accuracy and interpretability—allowing engineers and data scientists to understand exactly how the model identifies faults at different abstraction levels. This transparency is crucial for building trust in high-stakes applications like predictive maintenance, where incorrect diagnoses can lead to equipment failure or safety hazards.

Timeline & Sources

Jun 16, 2026

Wire

RGNet paper submitted to arXiv

Jun 18, 2026

Wire

RGNet paper published on arXiv

Entities

Sources

Neural Network Implementation of the Renormalization Group for Fault Diagnosis with Class Imbalancearxiv_csMediaJun 18, 2026