Balanced Adaptive Logit-Compensated Cross-Entropy and Quadratic Convolutional Network for Intelligent Fault Diagnosis...

Area of Science:

• Machine Learning
• Industrial AI
• Deep Learning for Diagnostics

Background:

• Long-tailed data distributions are prevalent in industrial settings, posing challenges for deep learning models due to performance degradation and bias.
• Existing methods struggle with imbalanced datasets, particularly in identifying rare equipment failure events.

Purpose of the Study:

• To propose a novel network, BALQNet, to enhance diagnostic performance on long-tailed data.
• To address the limitations of current deep learning approaches in handling imbalanced industrial datasets.

Main Methods:

• Developed BALQNet, a network combining balanced adaptive logit-compensated cross-entropy loss (BAL) and a quadratic convolution backbone.
• BAL integrates logit compensation, label smoothing, and class reweighting to optimize minority class learning.
• Quadratic convolution enhances feature representation learning.

Main Results:

• BALQNet demonstrated strong diagnostic performance on imbalanced bearing, gear, and motor datasets.
• The proposed method effectively handles long-tailed data distributions without significant performance degradation.
• Ablation studies confirmed the effectiveness of the BALQNet approach.

Conclusions:

• BALQNet offers a robust solution for improving deep learning-based diagnostics in industrial scenarios with imbalanced data.
• The combined approach of BAL and quadratic convolution effectively addresses the challenges posed by long-tailed distributions.
• The findings suggest BALQNet can significantly improve the reliability of industrial equipment monitoring systems.