This paper deals with the online model identification of robots suffering from backlash in their transmission, such as is the case for cable-driven endoscopic robots. Although online backlash identification is advantageous due to potential in-situ evolution of the backlash behavior, most existing methods focus on offline identification. Existing online approaches are either designed for DOF-by-DOF identification or limited by the simplicity of the underlying backlash models. We propose a new identification method based on Discontinuous EKF (DEKF) filtering to learn, online, the correct parameters of a multi-DOF backlash model. This allows to account for measurement noise and to include more complex backlash behaviors. A proof of concept on a simulated 3 DOF flexible endoscopic robot is presented to demonstrate the potential of such an approach.