A Learning-Based Approach for Diagnosis and Diagnosability of Unknown Discrete Event Systems

Area of Science:

• Control Engineering
• Computer Science
• Systems Theory

Background:

• Fault diagnosis in discrete event systems (DES) is crucial for system reliability.
• Existing methods may require complete system models or extensive prior knowledge.
• Unknown system dynamics pose a significant challenge for effective fault detection.

Purpose of the Study:

• To develop a novel active-learning technique for fault diagnosis of initially unknown finite-state DES.
• To construct a diagnoser capable of detecting and identifying faults by observing system behavior.
• To ensure the developed diagnoser is minimal and the fault diagnosis process is bounded.

Main Methods:

• Utilizing an active-learning mechanism to incrementally gather system information.
• Systematically completing observation tables to construct the diagnoser.
• Developing a deterministic finite-state automaton as the diagnoser.

Main Results:

• The proposed algorithm guarantees termination after a finite number of iterations.
• A correctly conjectured, minimal state diagnoser is produced.
• A sufficient condition for diagnosability is derived, ensuring bounded fault diagnosis.
• The method demonstrated capability in diagnosing multiple faults in case studies.

Conclusions:

• The active-learning approach provides an effective method for fault diagnosis in unknown DES.
• The developed diagnoser is proven to be minimal and accurate.
• The technique offers a robust solution for identifying system faults within a bounded number of observations.