Deep learning-based open set multi-source domain adaptation with complementary transferability metric for mechanical...

Area of Science:

• Mechanical Engineering
• Artificial Intelligence
• Machine Learning

Background:

• Intelligent fault diagnosis models struggle with inconsistent data distributions (domain shift) and the emergence of unknown fault modes not present in training data (category gap).
• These challenges limit the robustness and reliability of mechanical condition recognition systems.

Purpose of the Study:

• To develop an advanced open set multi-source domain adaptation approach to overcome domain shift and category gaps in intelligent fault diagnosis.
• To enhance the performance of mechanical condition recognition models in the presence of novel, unobserved fault types.

Main Methods:

• A complementary transferability metric was introduced to quantify target sample similarity to known classes, weighting an adversarial mechanism.
• An unknown mode detector was implemented for automatic identification of novel fault modes.
• A multi-source mutual-supervised strategy was employed to leverage information across different data sources.

Main Results:

• The proposed method demonstrated superior performance compared to traditional domain adaptation techniques on rotating machinery datasets.
• The approach effectively handled both domain shift and the challenge of identifying previously unseen fault modes.
• Experimental validation confirmed the robustness and accuracy of the developed fault diagnosis system.

Conclusions:

• The developed open set multi-source domain adaptation approach significantly improves intelligent fault diagnosis capabilities.
• This method offers a robust solution for real-world mechanical systems where data variability and unknown faults are common.
• The findings pave the way for more reliable and adaptable condition monitoring systems in mechanical engineering.