Failure Identification Using Model-Implemented Fault Injection with Domain Knowledge-Guided Reinforcement Learning.

Area of Science:

• Cyber-Physical Systems (CPS) Safety Engineering
• Automated Software Testing
• Machine Learning Applications

Background:

• Increasing complexity of cyber-physical systems (CPS) necessitates advanced safety assessment techniques.
• Traditional fault injection (FI) methods for finding catastrophic faults are labor-intensive, costly, and often inefficient.
• Identifying rare but critical faults early in development is crucial for robust system safety.

Purpose of the Study:

• To propose and evaluate a reinforcement learning (RL)-based method for automated fault configuration and detection in CPS.
• To leverage system (safety) specifications to guide the RL agent's learning process for effective fault identification.
• To demonstrate the method's capability in finding severe, catastrophic faults at the model level during early development stages.

Main Methods:

• Development of a reinforcement learning (RL) agent trained to automatically configure faults within a system model.
• Integration of system safety specifications into the RL agent's reward function to prioritize critical fault discovery.
• Dynamic interaction between the RL agent and the model under test (MATLAB/Simulink) to identify catastrophic faults.

Main Results:

• The proposed RL-based fault injection method successfully identified catastrophic faults in CPS models.
• Comparison with random-based FI in two case studies showed the RL method's superiority.
• The RL approach found a greater number and more severe faults compared to random fault injection.

Conclusions:

• Reinforcement learning offers an effective and automated approach for fault injection in cyber-physical systems.
• The RL-based method enhances the efficiency and effectiveness of discovering critical system faults early in development.
• This technique provides a valuable guideline for utilizing domain knowledge in RL-driven safety assessments of complex systems.