Subsystem-Based Fault Detection in Robotics via L2 Norm and Random Forest Models

Robotic systems require effective fault detection, achievable through various methods, including data-driven techniques. This paper proposes a novel fault detection method that leverages the
$L2$
norm of sensor measurements across different subsystems. Our approach involves dividing the system into subsystems, processing signals from a single sensor type, and using optimal features derived from the
$L2$
norm for prediction via a random forest model for detection. This single-signal approach allows the algorithm to accurately classify data without being distracted by different types of signals, in addition to eliminating the curse of dimensionality. The approach was validated on the Robot Execution Failures and Voraus-AD datasets, achieving exceptional performance across various metrics while significantly reducing data storage requirements, demonstrating both cost-effectiveness and computational efficiency.