Digital Twin for Training Bayesian Networks for Fault Diagnostics of Manufacturing Systems

Area of Science:

• Smart Manufacturing
• Industrial IoT
• Cyber-Physical Systems

Background:

• Smart manufacturing systems require resilience through rapid fault diagnosis.
• Existing methods often need extensive, balanced datasets from physical systems.
• Digital twins (DTs) offer a virtual environment for system development and testing.

Purpose of the Study:

• To propose a co-simulation approach for engineering DTs to train Bayesian Networks (BNs) for fault diagnostics.
• To enable fault injection in a virtual environment, bypassing the need for physical experimentation and balanced data.
• To introduce a Structural Intervention Algorithm (SIA) for accurate BN structure learning.

Main Methods:

• Engineered a co-simulation model using cyber-physical systems (CPS), high-fidelity equipment models, and discrete-event simulation (DES) of a factory.
• Developed a DT research test-bed with four industrial robots and IoT sensors.
• Integrated equipment simulators with a DES model for a robotic assembly cell.
• Employed the Structural Intervention Algorithm (SIA) for BN structure learning.

Main Results:

• The DT approach successfully enabled fault injection in the virtual system.
• The SIA algorithm accurately detected directed edges and distinguished parent/ancestor nodes in BNs.
• Laboratory experiments validated the approach, achieving significantly better accuracy (Structural Hamming Distance) than traditional methods.
• The learned BN structure demonstrated robustness against parameter variations like mean time to failure (MTTF).

Conclusions:

• The proposed co-simulation DT approach effectively trains BNs for equipment and factory-level fault diagnostics in smart manufacturing.
• This method alleviates the need for costly physical experiments and balanced datasets.
• The SIA enhances BN structure learning accuracy and robustness for improved performance assurance.