Adhesion Loss Prediction of a Climbing Robot through Magnetic Field Analysis by Artificial Neural Networks

Abstract. This paper discusses an improvement for an autonomous robot’s magnetic climbing system necessary to perform the inspection of metal plates and weld beads in internal/external LPG storage sphere’s surfaces. This task typically covers up to 360◦ in the roll-pitch-yaw angles navigation with respect to the Earth‘s surface plan. Also, the storage spheres’ metal surfaces present non-modeled disturbances (such as surface irregularities, rust dust, welding seams...) that can change the adhesion force, requiring an active control to maintain the force balance needed for trajectory tracking. Classic approaches utilize distance sensors to detect gaps between the wheel and the surface, providing a low-level feedback to the control system that, generally, does not have an active compensation of adhesion, thus, disturbing the robot navigation. Unfortunately this method is inappropriate due to the adhesion system’s nonlinear behavior (adhesion force between wheel and surface decreases radically with gaps) and tracking limitations. In order to surmount these problems, an adhesion disturbances recognition system based in artificial neural networks is developed to predict adherence anomalies through magnetic field analysis. This prediction of adherence loss allows the implementation of a new approach to adhesion feedback in robot navigation controllers, that could be used to overcome adhesion disturbances. The proposed system fulfills the inspection tasks by preventing the detachment of one or more wheels and, possibly, the robot’s fall without making path deviations.