Few-shot RUL prediction for engines based on CNN-GRU model

Area of Science:

• Prognostics and Health Management (PHM)
• Machine Learning for Engineering Systems

Background:

• Predicting Remaining Useful Life (RUL) for critical components like aircraft engines is vital for operational safety and maintenance.
• Insufficient historical life data poses significant challenges in feature extraction, temporal relationship modeling, and accurate RUL prediction.

Purpose of the Study:

• To propose a novel 1D Convolutional Neural Network (CNN) and Gated Recurrent Unit (GRU) hybrid model for RUL prediction under few-shot learning conditions.
• To enhance the accuracy and robustness of RUL prognostication by effectively utilizing limited historical data.

Main Methods:

• A 1D CNN-GRU hybrid network was developed, leveraging CNN for high-dimensional feature extraction and GRU for temporal feature analysis.
• Data distribution, correlation, and monotonicity indices were integrated to refine multi-sensor input parameters for the CNN-GRU model.
• The model was trained and validated using a sub-dataset from the NASA C-MAPSS multi-constraint dataset.

Main Results:

• The proposed 1D CNN-GRU model demonstrated high accuracy in RUL prediction tasks.
• Experimental results validated the model's effectiveness in handling few-shot learning scenarios for engine prognostics.
• The method successfully addressed challenges related to insufficient data for performance degradation feature extraction and temporal relationship modeling.

Conclusions:

• The developed CNN-GRU hybrid model offers a powerful and accurate solution for RUL prediction, particularly in data-scarce environments.
• This approach significantly improves predictive accuracy by combining deep learning architectures for feature and temporal extraction.
• The study highlights the potential of advanced machine learning techniques in enhancing the reliability and safety of critical engineering systems.