A Neural Network-Based Approach for the Identification and Compensation of Magnetic Disturbances in Mobile Robot Localization

Abstract

Magnetometers are widely used in sensor fusion frameworks of mobile robots due to their ability to work as compasses. The measurements of these sensors are largely influenced by hard iron and soft iron effects that are caused by nearby ferromagnetic objects. In this paper, an artificial neural network (ANN)-based method is proposed for the identification of such effects, which can be compensated from the measurements. The method utilizes both the distance and the angle of the detected object based on the coordinate frame of the mobile robot as the inputs of the Multi-Layer Perceptron (MLP) ANN. The MLP provides on its outputs the disturbance levels in the three axes, which need to be subtracted from the measurements to obtain the compensated values. The method was tested with three different types of objects to evaluate the performance of the method. Measurements were collected using an industrial robotic arm in a grid around the objects. The disturbance levels generated by the objects were computed by subtracting measurements recorded in an undisturbed scenario. Three-layer MLPs were tested with various numbers of hidden layer neurons to find the optimal configurations. The achieved results show that the disturbance levels can be significantly decreased using the proposed method.