Sim-to-Real Fault Detection and Predictive Maintenance for Collaborative Robots: Controller-Residual Feature Fusion and Domain-Scale Correction on the Universal Robots UR5e

This dissertation develops a real-time fault detection and predictive maintenance system for the Universal Robots UR5e collaborative robot, addressing a gap in existing fault-detection literature: no published work combines controller-residual signals with deep learning for collaborative robot diagnostics. The core contribution is the controller-residual tracking error feature (), evaluated alongside conventional sensor features across a combined simulation-and-real dataset (164,821 feature windows: 114,766 simulated, 50,055 from physical UR5 hardware) spanning four fault types at three severities plus normal operation.