Physics-Informed Monotonic Conformer for Remaining Useful Life Prediction of Hydraulic Systems

Area of Science:

• Mechanical Engineering
• Artificial Intelligence
• Reliability Engineering

Background:

• Hydraulic systems in heavy machinery require accurate health monitoring.
• Current data-driven models fail to capture long-term degradation and physical laws.
• Ignoring physical constraints leads to unrealistic fault progression predictions.

Purpose of the Study:

• To develop a novel model that integrates physical laws into data-driven health assessments.
• To improve the tracking of long-term degradation trends in hydraulic systems.
• To enhance the physical consistency of equipment fault predictions.

Main Methods:

• Introduction of the Physics-Informed Monotonic Conformer (PIMC).
• Combination of convolutional inductive biases and global self-attention for feature extraction.
• Implementation of a monotonicity loss function to enforce physical degradation constraints.

Main Results:

• The PIMC model outperforms existing baseline models on an electrohydrostatic actuator dataset.
• The monotonicity loss significantly improves the physical consistency of predictions.
• Achieved a high Spearman rank correlation, indicating reliable health indicators.

Conclusions:

• The PIMC model provides numerically precise and physically reliable health indicators for hydraulic systems.
• This approach grounds data-driven models in physical realities, overcoming limitations of purely data-driven methods.
• The developed health indicators are suitable for safety-critical applications like aerospace.