Remaining Useful Life Estimation for Engineered Systems Operating under Uncertainty with Causal GraphNets.

Area of Science:

• Machine Learning
• Reliability Engineering
• Signal Processing

Background:

• Predicting Remaining Useful Life (RUL) is crucial for proactive maintenance.
• Time-series data often presents challenges like non-equidistant observations and multiple temporal scales.
• Existing models may struggle with irregularly sampled data and implicit degradation patterns.

Purpose of the Study:

• To introduce a novel GNN-tCNN approach for constructing and training RUL models.
• To address the challenge of learning from time-series data with non-equidistant observations.
• To develop a method for representing prediction uncertainty in RUL estimations.

Main Methods:

• Utilizing Graph Neural Networks (GNNs) and message-passing neural networks.
• Encoding irregularly sampled causal structures within the time-series data.
• Employing reparametrization gradients to represent RUL prediction uncertainty as a gamma distribution.

Main Results:

• Demonstrated efficacy on both simulated stochastic degradation and real-world ball-bearing accelerated life testing datasets.
• The GNN-tCNN model implicitly learns system evolution, outperforming a comparable LSTM-tCNN approach.
• Successfully represented prediction uncertainty using a gamma distribution.

Conclusions:

• The GNN-tCNN approach offers an effective solution for RUL prediction with challenging time-series data.
• Implicitly modeling system degradation provides a robust alternative to raw observation-level analysis.
• Accurate uncertainty quantification enhances the reliability of RUL predictions.