Fault diagnosis through system-level condition-monitoring and digital-twin-supported deep learning

Diagnosing component-level faults in complex systems is a challenging task, particularly when only systemlevel condition-monitoring data are available due to difficulties in deploying sensors at the component level.

To address this issue, we propose a digital twin-supported deep learning framework for diagnosing componentlevel failures using system-level condition-monitoring data. The framework operates in two phases: an offline phase, where a Digital Failure Twin (DFT) model simulates failure modes and generates synthetic training data, and an online phase, where a sim-to-real error correction mechanism aligns simulation outputs with real-world system behavior using an error simulator and hyper-parameter tuning. This alignment ensures the diagnostic model effectively bridges the gap between simulated and real-world conditions. The proposed framework is evaluated on a real-world robot system where only the movement trajectory of the end-effector is used as condition-monitoring data to diagnose the failure modes of the four motors of the robot. An open-source DFT model for robotic systems was developed to generate synthetic failure data for training the diagnosis models, while real-world test data were collected to assess the model's performance. Experimental results demonstrate the effectiveness of the developed framework: it improves fault diagnosis accuracy on real datasets by up to 403.76% as compared to traditional deep learning models without using the synthetic data from the DFT, especially in scenarios with limited real data. Furthermore, with only 10 real operational samples, the developed sim-to-real correction method improves accuracy on the real test data by 15.51%.