Explainable Fault Diagnosis Using Invertible Neural Networks-A Left Manifold-Based Solution.

Area of Science:

• Control Engineering
• Machine Learning
• System Identification

Background:

• Nonlinear feedback control systems present challenges for traditional fault diagnosis (FD) methods.
• Existing FD approaches may struggle with overfitting when learning complex nonlinear dynamics.
• There is a need for interpretable and specialized FD techniques for these systems.

Purpose of the Study:

• To develop a novel intelligent fault diagnosis (FD) paradigm for nonlinear feedback control systems.
• To design an Invertible Neural Network (INN)-based FD scheme utilizing a left manifold.
• To enhance system identification accuracy and avoid the overfitting problem in nonlinear dynamics.

Main Methods:

• Formulation of a residual generator as a projection of system data onto a null space.
• Elaboration of a homeomorphism in a topological space for an invertible relationship between system outputs and residuals.
• Introduction of master and slave objective functions for information-lossless system/parameter identification.

Main Results:

• Demonstrated the feasibility of the Invertible Left Manifold (ILM)-based FD strategy through two nonlinear system studies.
• The proposed INN-based FD scheme effectively avoids the overfitting problem in nonlinear system dynamics.
• The control theory-guided design ensures interpretability of the learning process.

Conclusions:

• The developed INN-based FD paradigm offers a specialized and effective solution for nonlinear feedback control systems.
• This research contributes to advancements in machine learning-based system identification and explainable FD.
• The findings pave the way for future research, including right manifold-based FD designs in Part II.