Calibrated Adaptive Teacher for Domain-Adaptive Intelligent Fault Diagnosis.

Area of Science:

• Engineering
• Computer Science
• Machine Learning

Background:

• Deep learning for intelligent fault diagnosis (IFD) excels with raw signals but struggles with distribution shifts.
• Unsupervised domain adaptation (UDA) addresses IFD when labeled source data and unlabeled target data are available.
• Existing UDA methods using pseudo-labeling are limited by poor uncertainty calibration and over-confident predictions.

Purpose of the Study:

• To propose a novel method, Calibrated Adaptive Teacher (CAT), for domain-adaptive IFD.
• To address the challenge of poorly calibrated uncertainty estimates in target domains during self-training.
• To enhance the quality of pseudo-labels and mitigate error accumulation in UDA for IFD.

Main Methods:

• Developed the Calibrated Adaptive Teacher (CAT) framework for UDA in IFD.
• Integrated post hoc calibration techniques to calibrate teacher network predictions on target samples.
• Evaluated CAT using temperature scaling, one of four tested calibration methods, on the Paderborn University (PU) benchmark.
• Utilized both time- and frequency-domain inputs for rolling bearing fault diagnosis under varying conditions.

Main Results:

• The CAT method, specifically CAT+TempScaling, achieved state-of-the-art performance on most transfer tasks.
• Demonstrated an average accuracy increase of 7.5% compared to domain-adversarial neural networks (DANNs).
• Achieved a 4 times lower calibration error compared to DANNs across twelve transfer tasks.

Conclusions:

• Calibrating teacher network predictions is crucial for effective self-training in domain-adaptive IFD.
• CAT, particularly with temperature scaling, significantly improves accuracy and reduces calibration error in challenging transfer learning scenarios.
• The proposed method offers a robust solution for deploying IFD models in real-world conditions with varying operational distributions.