Dynamically Weighted Spatiotemporal Fusion for Deep Learning-Based Prediction of EHA Degradation in Aviation Systems

Area of Science:

• Aerospace Engineering
• Machine Learning
• System Health Monitoring

Background:

• Electro-hydrostatic actuators (EHAs) are critical in modern aircraft but pose challenges for health monitoring.
• Existing data lacks explicit links to system health, hindering degradation assessment and Remaining Useful Life (RUL) prediction.

Purpose of the Study:

• To develop a spatiotemporal degradation modeling framework (PreDyn-ST) for EHAs using multivariate time series data.
• To improve the accuracy and interpretability of degradation assessment and RUL prediction for aerospace systems.

Main Methods:

• Utilizes SimCLR-based contrastive pretraining and a dynamic feature fusion mechanism.
• Employs Graph Convolutional Networks (GCNs) for spatial modeling and Transformers for temporal pattern extraction.
• Incorporates a learnable dynamic weighting mechanism for adaptive feature balancing and analyzes interpretability using Correlation Statistical Index (CSI) curves.

Main Results:

• PreDyn-ST demonstrates competitive and stable prediction performance on EHA degradation data and the C-MAPSS benchmark.
• The framework shows robust performance under complex operating conditions, including the FD004 dataset.
• Achieves accurate and interpretable degradation modeling for aerospace applications.

Conclusions:

• The proposed PreDyn-ST framework effectively addresses the limitations of current EHA health monitoring.
• It provides a robust and interpretable solution for accurate degradation modeling and RUL prediction in aerospace.
• Highlights the potential for advanced machine learning techniques in enhancing aircraft safety and maintenance.