Data-driven remaining useful life prediction based on domain adaptation.

Area of Science:

• Prognostics and Health Management (PHM)
• Machine Learning for Predictive Maintenance

Background:

• Remaining Useful Life (RUL) prediction is crucial for PHM, enhancing system reliability.
• Data-driven methods excel in RUL prediction but typically require extensive labeled data and assume similar data distributions.
• Real-world data variability due to operating conditions, faults, and noise often violates this distribution assumption.

Purpose of the Study:

• To propose a novel data-driven framework for RUL prediction that incorporates domain adaptability.
• To overcome limitations of traditional methods requiring identical data distributions for training and testing.
• To improve the generalization capability of RUL prediction models in the face of data heterogeneity.

Main Methods:

• A data-driven framework utilizing a bidirectional gated recurrent unit (BGRU) was developed.
• Domain-adversarial neural network (DANN) was employed to implement transfer learning (TL) from a source to a target domain.
• The framework was designed to handle sensor data exclusively in the target domain.

Main Results:

• The proposed framework demonstrated improved generalization ability on the IEEE PHM 2012 Challenge datasets.
• Domain adaptation effectively mitigated performance degradation caused by data distribution discrepancies.
• The BGRU-based model with DANN achieved robust RUL prediction performance.

Conclusions:

• The proposed domain-adaptive framework effectively enhances RUL prediction accuracy and reliability.
• Transfer learning via DANN is a viable strategy for adapting models to different data distributions in PHM.
• This approach offers a promising solution for real-world RUL prediction challenges with limited or varied data.