Fault Decoupling Through State Estimation Using Observer-Based Physics-Informed Neural Network

Area of Science:

• Control Systems Engineering
• Fault Diagnosis
• Smart Manufacturing

Background:

• Fault decoupling is crucial for anomaly detection in smart manufacturing.
• Distinguishing between actuator faults, sensor faults, and external disturbances is challenging.

Purpose of the Study:

• To propose a joint state and fault estimation method for linear time-invariant (LTI) systems.
• To achieve mutual decoupling between multiple faults and disturbances.
• To enhance estimation accuracy by integrating physical knowledge.

Main Methods:

• Design of dual proportional-integral fading unknown input observers (PIFUIOs) for state and fault estimation.
• Development of a physics-informed neural network (PINN) using observer-based loss functions.
• Formulation of sufficient conditions using Lyapunov stability theory and linear matrix inequalities (LMIs).

Main Results:

• Guaranteed existence of observers and mutual decoupling between faults and disturbances.
• Enhanced estimation accuracy by incorporating physical knowledge into the PINN.
• Demonstrated asymptotic convergence and attenuation of faults and disturbances.

Conclusions:

• The proposed joint state and fault estimation approach effectively addresses fault diagnosis in LTI systems.
• The integration of PIFUIOs and PINNs offers a robust solution for smart manufacturing.
• Validation on a DC motor model confirms the approach's practical effectiveness.