Transfer learning and domain adaptation play a crucial role in enabling predictive models to generalize across different machines, operating conditions, and data distributions. In industrial applications, domain shifts between the training and target environments often lead to significant performance degradation, making it essential to develop methods that ensure robust adaptation.
Domain shifts can be categorized as:
- Continuous Shifts: Gradual changes in operating conditions, such as varying temperatures or load levels over time, or shifts due to maintenance and operational adjustments.
- Discrete Shifts: Significant changes occurring between different domains, such as from simulation to real-world environments or between different units of a fleet operating under distinct conditions.
Approaches to Transfer Learning and Domain Adaptation
Our research explores various techniques to improve knowledge transfer across domains, with a strong emphasis on integrating expert knowledge, including:
- Synthetic-to-Real Transfer Learning: Using synthetic data augmented with expert knowledge to enhance model performance when real-world fault data is scarce [1].
- Phase-Aware Domain Adaptation: Incorporating operational phases into the adaptation process to address phase misalignment issues in RUL prediction [2].
- Graph-Based Domain Adaptation: Applying graph neural networks to capture spatial-temporal dependencies and align sensor data distributions between domains [3].
- Uncertainty-Guided Adaptation: Incorporating uncertainty quantification techniques to guide adaptation and mitigate the impact of domain shifts in regression tasks (important for prognostics) [4,5].
- Test-Time Adaptation: Enabling real-time adaptation to evolving operating conditions without the need for additional labeled data [5].
By leveraging these approaches and integrating expert knowledge, our research aims to develop practical and scalable solutions that facilitate reliable model deployment across different industrial environments.
Papers
[1] Wang, Q., Taal, C., & Fink, O. (2021). Integrating expert knowledge with domain adaptation for unsupervised fault diagnosis. IEEE Transactions on Instrumentation and Measurement, 71, 1-12.
[2] Nejjar, I., Geissmann, F., Zhao, M., Taal, C., & Fink, O. (2024). Domain adaptation via alignment of operation profile for Remaining Useful Lifetime prediction. Reliability Engineering & System Safety, 242, 109718.
[3] Zhang, Z., & Fink, O. (2024). Domain Adaptive Unfolded Graph Neural Networks. arXiv preprint arXiv:2411.13137.
[4] Nejjar, I., Frusque, G., Forest, F., & Fink, O. (2024). Uncertainty-Guided Alignment for Unsupervised Domain Adaptation in Regression. arXiv preprint arXiv:2401.13721.
[5] Faghih Niresi, K., Nejjar, I., & Fink, O. (2024). Efficient Unsupervised Domain Adaptation Regression for Spatial-Temporal Air Quality Sensor Fusion. arXiv e-prints, arXiv-2411.
[6] Sun, H., Ammann, K., Giannoulakis, S., & Fink, O. (2024). Continuous Test-time Domain Adaptation for Efficient Fault Detection under Evolving Operating Conditions. arXiv preprint arXiv:2406.06607.