GNSS/MEMS INS tightly coupled algorithm for agricultural machinery navigation enhanced by random forest-based behavioral state awareness

Abstract

Global navigation satellite system (GNSS)/micro-electro-mechanical systems inertial navigation system (MEMS INS) algorithm is widely used in agricultural machinery navigation. However, several issues remain noteworthy, uneven terrain induces more high-frequency noise than other environments, which severely affects the accuracy of MEMS INS. In addition, occlusion environment, such as field windbreak, degrades GNSS signal quality. Although Butterworth filter and non-holonomic constraints (NHC) have been validated as effective solutions for these issues, which still face the following limitations in agricultural scenarios. This is because the power spectral density (PSD) of MEMS INS data exhibits distinct energy distribution among different states, therefore it is unreasonable to apply uniform cutoff frequency same as classic Butterworth filter. Additionally, jumping and slipping frequently occur, which can invalidate the zero-velocity assumption of NHC. Therefore, given the limitations of previous studies, this paper proposes a random forest (RF)-based model to identify machinery states and predict body-frame (right and up) velocities. Then, adaptive cutoff frequencies are selected for the Butterworth filter. Furthermore, the measurement and stochastic models of NHC are optimized by states and body-frame velocities. Experiments show that the proposed algorithm can achieve centimeter-level positioning accuracy and the heading angle error of only 0.33°.