Intelligent fault diagnosis scheme via multi-module supervised-learning network with essential features...

Area of Science:

• Machine Learning
• Artificial Intelligence
• Signal Processing

Background:

• Intelligent diagnosis methods rely heavily on data quantity and quality.
• Acquiring sufficient failure data for training is a significant challenge.
• Limited data leads to unsatisfactory model training and performance degradation.

Purpose of the Study:

• To address the challenge of insufficient failure data in intelligent diagnosis.
• To propose a novel method for constructing artificial samples to augment fault data volume.
• To improve the performance and accuracy of intelligent diagnostic models under small sample conditions.

Main Methods:

• A Signal Adaptive Augmentation Network (SAAN) was developed, comprising an impulse extractor, regulator, and classifier.
• The impulse extractor uses inner product matching to identify local impulse features for initial sample generation.
• The regulator employs convolution and deconvolution with a synthetic loss function to ensure artificial samples match real data distributions.

Main Results:

• The SAAN method was validated on three bearing datasets against advanced algorithms.
• A focal normalized network was utilized for classification with small sample sizes.
• Experiments demonstrated that SAAN improves recognition accuracies of diagnostic models by 5%-35%.

Conclusions:

• The proposed SAAN effectively amplifies fault data volume by creating realistic artificial samples.
• SAAN offers a competitive and effective solution for data-driven intelligent diagnosis with limited data.
• The method significantly enhances the accuracy of diagnostic models in small sample scenarios.