Max-product algorithms for the generalized multiple-fault diagnosis problem

Area of Science:

• Computer Science
• Electrical Engineering
• Systems Engineering

Background:

• The generalized multiple-fault diagnosis (GMFD) problem involves identifying faults in systems with unreliable components and alarms.
• Existing diagnostic methods face challenges in complex systems with interconnected components and potential cycles.

Purpose of the Study:

• To apply and evaluate the max-product algorithm (MPA) and a novel sequential MPA (SMPA) for GMFD.
• To analyze algorithm performance on bipartite diagnosis graphs (BDGs), including acyclic and cyclic cases.

Main Methods:

• Developed and applied MPA and SMPA, which are local message-passing algorithms.
• Utilized bipartite diagnosis graphs (BDGs) to model the GMFD problem.
• Conducted simulations and analytical assessments for various graph structures and alarm patterns.

Main Results:

• MPA and SMPA converge to the maximum a posteriori probability (MAP) solution for acyclic BDGs.
• Simulations indicate strong performance for both MPA and SMPA in general systems, with SMPA outperforming MPA in systems with cycles.
• Analytical bounds on erroneous diagnosis probability were derived for general graphs.

Conclusions:

• MPA and SMPA offer efficient solutions for GMFD with reasonable computational complexity, particularly for sparse diagnosis graphs.
• SMPA demonstrates enhanced performance over MPA in more complex, cyclic system models.
• The algorithms provide a robust approach to fault diagnosis in unreliable systems.