A Spatio-Temporal Attention Mechanism Based Approach for Remaining Useful Life Prediction of Turbofan Engine

Area of Science:

• Aerospace Engineering
• Machine Learning
• Predictive Maintenance

Background:

• Turbofan engine time-series data exhibit high complexity and dynamics.
• Recurrent Neural Networks (RNNs) are standard for Remaining Useful Life (RUL) prediction but struggle with data relationships and gradient issues.
• Existing methods often overlook spatial-temporal dependencies in turbofan engine data.

Purpose of the Study:

• To propose a novel spatio-temporal attention model for improved RUL prediction.
• To address limitations of RNNs by incorporating both temporal and spatial data feature relationships.
• To enhance the stability and accuracy of RUL forecasting in turbofan engines.

Main Methods:

• Developed a spatio-temporal attention model integrating temporal data associations and spatial hidden states.
• Implemented positional encoding for temporal relationships, bypassing the need for recurrent neural networks.
• Validated the model on the Commercial Modular Aerospace Propulsion System Simulation (C-MAPSS) datasets.

Main Results:

• The spatio-temporal attention model significantly improved predictive performance by leveraging combined temporal and spatial dimensions.
• Experimental results demonstrated superior stability and prediction accuracy compared to alternative methods on C-MAPSS datasets.
• The proposed model effectively captures complex data dynamics for more reliable RUL forecasting.

Conclusions:

• The spatio-temporal attention model offers a robust alternative to traditional RNNs for turbofan engine RUL prediction.
• Integrating spatio-temporal features enhances model performance, providing more accurate and stable prognostics.
• This approach advances predictive maintenance strategies in aerospace applications.