Practical Test-Time Domain Adaptation for Industrial Condition Monitoring by Leveraging Normal-Class Data.

Area of Science:

• Machine Learning
• Industrial IoT
• Data Science

Background:

• Machine learning models degrade with domain shift, impacting industrial condition monitoring.
• Existing domain adaptation methods are not suitable for real-world industrial settings.

Purpose of the Study:

• To develop a practical domain adaptation framework for industrial condition monitoring.
• To address performance degradation caused by distributional shifts in sensor data.

Main Methods:

• Introduced a Normal-Class Test-Time Domain Adaptation (NC-TTDA) framework.
• Framework detects shifts and adapts models using normal-class samples without labeled target data.
• Integrated with automated machine learning (AutoML) for end-to-end optimization.

Main Results:

• Achieved robust generalization across six condition monitoring datasets under domain shift.
• Attained average AUROC scores exceeding 99% with low false positive rates.
• Demonstrated effectiveness in real-world industrial monitoring scenarios.

Conclusions:

• NC-TTDA provides a practical solution for domain adaptation in condition monitoring.
• The framework enhances the reliability and generalizability of machine learning models in industrial applications.