Intermittent fault detection for MIMO systems: a case study on SCARA robot

Abstract

In the domain of robotics and automation, the accurate supervision of key parameters is essential to ensure the safe and efficient operation of various systems. This paper presents a comparative study for fault detection based on data-driven approaches to estimate actuator torques subjected to different kind of intermittent faults in a two-degree-of-freedom SCARA robot. Different supervised machine learning techniques are used: Gaussian Process Regression with a Matérn 5/2 kernel, Ensemble Random Forest Regression, Support Vector Machine with a medium Gaussian kernel, and a Narrow Neural Network. Two types of fault scenarios are defined including different evaluation criteria such computation time, R-squared, Mean Absolute Error, Mean Squared Error, and Root Mean Squared Error. The process of model selection involves not only these criteria but also factors like efficiency in both healthy and faulty scenarios, as well as the significant impact of a fault that occurred in one torque on the second torque. Furthermore, our selection is also based on the adequacy between the optimal computation time and model complexity. We conclude that, GPR with a Matérn 5/2 kernel function is the most appropriate one for the estimation compared with others for this type of application.