Predicting System Degradation with a Guided Neural Network Approach.

Area of Science:

• Engineering
• Materials Science
• Data Science

Background:

• Accurate lifetime estimation is vital for engineering system safety and reliability.
• Traditional methods like accelerated life testing have limitations in simulating real-world conditions.
• Field degradation measurement under actual operating conditions is often time-consuming.

Purpose of the Study:

• To develop a time-efficient, data-driven framework for modeling field degradation and predicting system lifetime.
• To address challenges in incorporating degradation physics and reducing extensive training data requirements for neural networks.
• To improve the accuracy and efficiency of lifetime estimations for engineering systems.

Main Methods:

• Proposed a framework integrating a physics discovery neural network and a predictive neural network.
• The physics discovery network models degradation physics, guiding the predictive network for enhanced life estimations.
• Utilized sensor measurements from short-term actual operating conditions degradation tests.

Main Results:

• Demonstrated effectiveness through a case study on atmospheric corrosion of steel in a marine environment.
• Achieved up to a 76% reduction in mean absolute error compared to standard neural network models.
• The integrated framework provided more accurate and efficient lifetime predictions.

Conclusions:

• The proposed data-driven framework enables efficient evaluation of system safety and reliability.
• Optimized maintenance activities through accurate and timely lifetime predictions.
• The approach successfully incorporates physical degradation principles into data-driven models for improved performance.