Sensor Fault Diagnostics Using Physics-Informed Transfer Learning Framework

Area of Science:

• Engineering
• Computer Science

Background:

• Intelligent sensor fault diagnostics are crucial for reliable operation in engineering applications.
• Identifying sensor fault types is a significant challenge in smart health monitoring.

Purpose of the Study:

• To extend a physics-informed transfer learning framework for sensor fault diagnostics.
• To apply this framework for the first time in identifying diverse sensor fault scenarios.

Main Methods:

• Utilized dynamic mode decomposition with control (DMDc) for data-driven extraction of system dynamics.
• Employed continuous wavelet transform (CWT) for generating time-frequency illustrations.
• Applied transfer learning with a pre-trained image classification algorithm for fault diagnosis.

Main Results:

• Achieved high performance in sensor fault diagnostics with 91.5% training accuracy and 84.7% test accuracy.
• Demonstrated a fair level of robustness across different fault scenarios (bias, drift, sine disturbance, increased noise).
• Successfully implemented intelligent sensor fault diagnostics using the extended framework.

Conclusions:

• The physics-informed transfer learning framework is effective for intelligent sensor fault diagnostics.
• This approach offers a promising solution for enhancing the reliability of smart health monitoring systems.