A Machine Learning Approach for an Improved Inertial Navigation System Solution

Abstract

The inertial navigation system (INS) is a basic component to obtain a continuous navigation solution in various applications. The INS suffers from a growing error over time. In particular, its navigation solution depends mainly on the quality and grade of the inertial measurement unit (IMU), which provides the INS with both accelerations and angular rates. However, low-cost small micro-electro-mechanical systems (MEMSs) suffer from huge error sources such as bias, the scale factor, scale factor instability, and highly non-linear noise. Therefore, MEMS-IMU measurements lead to drifts in the solutions when used as a control input to the INS. Accordingly, several approaches have been introduced to model and mitigate the errors associated with the IMU. In this paper, a machine-learning-based adaptive neuro-fuzzy inference system (ML-based-ANFIS) is proposed to leverage the performance of low-grade IMUs in two phases. The first phase was training 50% of the low-grade IMU measurements with a high-end IMU to generate a suitable error model. The second phase involved testing the developed model on the remaining low-grade IMU measurements. A real road trajectory was used to evaluate the performance of the proposed algorithm. The results showed the effectiveness of utilizing the proposed ML-ANFIS algorithm to remove the errors and improve the INS solution compared to the traditional one. An improvement of 70% in the 2D positioning and of 92% in the 2D velocity of the INS solution were attained when the proposed algorithm was applied compared to the traditional INS solution.

Generation of Correction Data for Autonomous Driving by Means of Machine Learning and On-Board Diagnostics

“…The research work that most resembles the methods proposed herein is shown in [ 20 ]. There, the authors train an Adaptive Neuro-Fuzzy Inference System (ANFIS) that receives six inertial measurements as an input—the angular velocities around and the accelerations along the x , y , and z axes of a consumer-grade Inertial Measurement Unit (IMU)—and the learning targets are the same inertial measurements but from a high-end IMU.…”

“…Their results show an improvement of 70% in the 2D positioning and 92% improvement in the velocity estimation, which shows that consumer-grade sensors can approach the measuring characteristics of their high-performance counterparts. The main differences between [ 20 ] and the present work are (1) the sensors used, (2) the ML-model applied, (3) the state variable to address, and (4) the analysis of the possibility for online implementation. With regards to (1), the present work does not use external sensors (as is the consumer-grade IMU), but only the on-board sensors of commercial vehicles instead.…”

A New Self-Calibration and Compensation Method for Installation Errors of Uniaxial Rotation Module Inertial Navigation System

“…Current calibration methods are usually based on laboratory turntables or external information-generated reference signals to calibrate installation errors in the SINS [ 28 ]. In recent years, to calibrate the error of inertial sensors accurately, some novel calibration methods based on machine learning have also been widely studied [ 29 , 30 ]. However, these methods generally use high-precision data to train a network model and then apply the model to low-precision data to improve performance.…”

Learning Car Speed Using Inertial Sensors

“…Previous attempts to apply supervised learning in estimation problems with localization applications were focused on learning the sensor noise models. Mahdi et al used low-cost sensors as input together with expensive sensors to generate the true labels for the measurement noise [12]. Azzam et al trained long short-term memory (LSTM) models to improve the trajectory accuracy obtained using a simultaneous localization and mapping (SLAM) algorithm [13].…”