Time-delayed multiple linear regression for increasing MEMS inertial sensor performance by using observations from a navigation-grade IMU

Abstract

Micro-electro mechanical systems (MEMS) inertial sensors are key components in navigation systems where low cost, low weight and/or low power consumption are required. New approaches based on machine learning techniques for nonlinear systems have been proposed to increase MEMS inertial sensors' precision. However, many MEMS inertial sensors can be considered almost linear according to the information provided by their manufacturers. In this work, a time-delayed multiple linear regression (TD-MLR) model is proposed to correct the non-deterministic sources of error of a MEMS inertial measurement unit (IMU). TD-MLR unknown coefficients are found by training a TD-MLR model with navigation-grade IMU observations. It is the authors' believe that this is a novel approach in applying machine learning techniques for improving the precision of MEMS inertial sensors. The generation and evaluation of TD-MLR models are achieved by using field data from a real trajectory. It is observed that, on average, the output of inertial sensors of different qualities is improved about 77% for accelerometers and 87% for gyroscopes by implementing the proposed machine learning procedure.