Temporal and Heterogeneous Graph Neural Network for Remaining Useful Life Prediction.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Predicting remaining useful life (RUL) is vital for industrial prognostics and health management.
• Deep learning models excel at identifying temporal dependencies in sensor data.
• Existing methods often miss fine-grained temporal information and sensor heterogeneity.

Purpose of the Study:

• To develop a novel model for RUL prediction that captures both temporal and spatial dependencies in sensor data.
• To leverage the heterogeneity of diverse sensor types for more accurate RUL predictions.
• To address the limitations of existing methods in capturing fine-grained temporal dynamics and sensor correlations.

Main Methods:

• Introduction of temporal and heterogeneous graph neural networks (THGNNs).
• THGNNs aggregate historical data from neighboring nodes for fine-grained temporal and spatial analysis.
• Feature-wise linear modulation (FiLM) is employed to handle sensor heterogeneity.

Main Results:

• THGNNs effectively capture fine-grained temporal dynamics and spatial correlations in sensor data.
• The model demonstrates significant improvements in RUL prediction accuracy.
• Achieved up to 19.2% and 31.6% improvement over state-of-the-art methods on the N-CMAPSS dataset.

Conclusions:

• THGNNs offer a powerful approach for RUL prediction by modeling complex sensor relationships.
• Leveraging sensor heterogeneity via FiLM enhances model performance.
• The proposed method represents a significant advancement in prognostics and health management.