HSELL-Net: A Heterogeneous Sample Enhancement Network With Lifelong Learning Under Industrial Small Samples

Area of Science:

• Engineering
• Computer Science

Background:

• Small sample sizes in industrial fault diagnosis lead to poor model accuracy and generalization.
• Traditional domain adaptation (DA) methods struggle with large variations in fault samples across different working conditions.

Purpose of the Study:

• To propose a novel network, HSELL-Net, for enhancing small industrial fault diagnosis samples.
• To improve the accuracy and generalization capabilities of fault diagnosis models with limited data.

Main Methods:

• A heterogeneous DA subnet (HDA-subnet) with dimension alignment and joint distribution matching to transfer knowledge from diverse domains.
• A lifelong learning subnet (LL-subnet) utilizing Admixup and a shared knowledge repository for incremental sample enhancement without retraining.
• Mutual embedding and reinforcement of the HDA-subnet and LL-subnet to increase sample diversity and quantity.

Main Results:

• HSELL-Net effectively enhances small fault diagnosis samples by leveraging heterogeneous domain data and incremental learning.
• The proposed method demonstrates superior performance compared to existing approaches in industrial small sample fault diagnosis scenarios.
• Experimental validation on both simulated and real-world data confirms the effectiveness of HSELL-Net.

Conclusions:

• HSELL-Net offers a robust solution for industrial fault diagnosis with limited data.
• The integration of heterogeneous domain adaptation and lifelong learning significantly improves model accuracy and generalization.
• The proposed approach addresses the limitations of traditional DA methods in handling variations across industrial working conditions.