Sliding mode based fault diagnosis with deep reinforcement learning add-ons for intrinsically redundant manipulators

This article presents a fault diagnosis control scheme for intrinsically redundant robot manipulators based on the combination of a deep reinforcement learning (DRL) approach and a battery of sliding mode observers. The DRL plays the role of detecting and isolating possible sensor faults, thus generating an alarm and pin-pointing the source. This in turn allows to compensate the sensor faults independently from the actuator ones. The latter are therefore detected and isolated by a set of sliding mode observers driven by input laws designed according to an optimal reaching algorithm. In order to design and apply such observers, a global feedback linearization is performed, which transforms the multi-input-multi-output (MIMO) nonlinear robot model into a chain of double integrators. The proposal is analyzed and assessed in realistic conditions using the PyBullet environment in which a 7 degrees-of-freedom (DOFs) Franka Emika Panda robot manipulator is reproduced.