A Comparison of Neural Networks and Fuzzy Inference Systems for the Identification of Magnetic Disturbances in Mobile Robot Localization

Abstract

Three-axis magnetometers are widely used in the field of localization in both outdoor and indoor environments. However, magnetic field measurements are disturbed by the presence of metallic objects due to the soft and hard-iron effects. To neglect these effects, a compensation technique is required, and, in this article, different solutions are proposed and evaluated to compensate for the disturbance effects of metallic objects with known fingerprints. These techniques exploit an already presented concept in the literature that is able to provide the compensation values of a known detected object using the distance and angle as inputs to a single hidden layer Artificial Neural Network (ANN). In this work, unlike the original proposal, each new presented technique exploits a modified or a different soft computing tool, such as a double hidden ANN, a Fuzzy Inference System (FIS), and an Adaptive Neural FIS (ANFIS). The techniques were tested with real measurements of three different objects, and the performances of the techniques were compared using the maximum errors, the Mean Absolute Errors (MAEs) of every single component, and the total MAEs. Overall, among them, only the ANN techniques and the ANFIS provided acceptable results. More precisely, the former provided maximum errors in the range between 0.3 μT and 3.8 μT, and MAEs in the order of 0.07 μT, whereas the latter was the one that provided the best performance, giving a residual maximum error in the order of 10 -3 μT and an MAE in the order of 10 -5 μT.