A Neural Network Approach to Mitigate Thermal-Induced Errors in ZUPT-aided INS

Abstract

Zero velocity UPdaTe (ZUPT)-aided Inertial Navigation Systems (INS) using foot-mounted Micro-Electro-Mechanical-Systems (MEMS) Inertial Measurement Units (IMUs) have been considered as a promising technology for localization of emergency responders, including firefighters and other personnel, in GPS-denied environments. Most commercially available general purpose MEMS IMUs are sensitive to ambient temperature changes. As a result, ZUPT-aided INS using these devices can have a degraded performance when operating in temperature-varying scenarios. This paper proposed a ZUPT-aided INS enhanced with a Back-Propagation Neural Network (BPNN)-based thermal compensation method. The proposed approach trained 12 different BPNNs to mitigate 12 thermal-induced errors separately, including bias drifts and noise standard deviation variations of accelerometers and gyroscopes along the three axes. We compared the proposed temperature-compensated ZUPT-aided INS with the traditional ZUPT-aided INS with a series of pedestrian indoor walking experiments in both temperature-static and -varying environments. Our experimental results showed that in the static cases, the traditional approach and our proposed approach had similar position Root-Mean-Squared Error (RMSE) of 0.38 m and 0.34 m, respectively. In the varying cases, however, the traditional approach had an RMSE of 9.29 m while our proposed approach significantly reduced the RMSE to 0.57 m.