Similarity-Based Remaining Useful Lifetime Prediction Method Considering Epistemic Uncertainty

Area of Science:

• Engineering
• Data Science
• Machine Learning

Background:

• Remaining Useful Life (RUL) prediction relies heavily on trajectory similarity measures.
• Existing methods often neglect epistemic uncertainty from asynchronous sampling or impose restrictive assumptions.
• These limitations lead to biased results and reduced prediction accuracy in RUL analysis.

Purpose of the Study:

• To develop a robust similarity measure for degradation trajectories that accounts for sampling uncertainty.
• To improve the accuracy and reliability of Remaining Useful Life (RUL) prediction models.
• To overcome the limitations of existing similarity metrics like EDR, LCSS, and DTW.

Main Methods:

• Proposed an uncertain ellipse model to represent sampling points as observations from an uncertain distribution.
• Developed a novel similarity measure metric for comparing degradation trajectories.
• Utilized a Stacked Denoising Autoencoder (SDA) for RUL prediction, incorporating similarity-based data selection and fine-tuning.

Main Results:

• The proposed similarity measure effectively models sampling uncertainty, unlike traditional methods.
• The SDA model trained on similar degradation data demonstrated superior RUL prediction performance.
• The new method showed greater robustness and accuracy compared to EDR, LCSS, and DTW across different sampling rates.

Conclusions:

• The novel uncertain ellipse-based similarity metric provides a more accurate and robust approach to trajectory comparison.
• The proposed RUL prediction framework, leveraging this metric and SDA, significantly outperforms existing methods.
• This work offers a promising direction for enhancing prognostics and health management in engineering systems.