Robust Calibration of MEMS Accelerometers in the Presence of Outliers

MEMS accelerometers suffer from different problems such as random and systematic errors and the presence of outliers in the measurements. Proper calibration is necessary to obtain accurate results. Various methods exist but very few are designed for robust calibration. This paper explores methods fo...

Full description

Saved in:
Bibliographic Details
Published in:IEEE sensors journal 2022-05, Vol.22 (10), p.9500-9508
Main Author: Belkhouche, Fethi
Format: Article
Language:English
Subjects:
Accelerometer calibration
Accelerometers
Algorithms
Anomaly detection
Calibration
Data analysis
Deviation
Ellipsoids
Linearization
Mathematical models
Median (statistics)
Microelectromechanical systems
Micromechanical devices
Optimization
Optimization techniques
outliers
Outliers (statistics)
robust estimation
Robustness
Sensors
Systematic errors
Citations: Items that this one cites
Items that cite this one
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:MEMS accelerometers suffer from different problems such as random and systematic errors and the presence of outliers in the measurements. Proper calibration is necessary to obtain accurate results. Various methods exist but very few are designed for robust calibration. This paper explores methods for calibrating MEMS accelerometers in the presence of outliers. Optimization techniques based on Levenberg-Marquardt algorithm and linearization are used to solve the nonlinear calibration equations. Outlier detection methods such as RANSAC (random sample consensus), the Mahalanobis distance, and the median absolute deviation (MAD) are integrated within the optimization algorithm. Outlier scores are calculated and used to eliminate outliers or assign weights. The experimental results show that outlier detection and removal methods allow to achieve substantial improvement in the calibration process compared with non-robust methods. The median absolute deviation is the most effective in detecting outliers. In the presence of outliers, the average error for MAD and model-based method is 0.0136 (a.u.). For non-robust methods, the error is 3.82 (a.u.). For the Mahalanobis distance, the error is 1.8778 (a.u.). The results show that the calibration error is the smallest for the model-based method. The Mahalanobis distance presents higher calibration errors and is less capable of detecting outliers; however, it is still better than non-robust methods. Finally Levenberg-Marquardt algorithm presents slightly better results than linearization.
ISSN:1530-437X
1558-1748
DOI:10.1109/JSEN.2022.3163964