Adaptive GNSS/INS Integration Based on Supervised Machine Learning Approach

Guohao Zhang, Li-Ta Hsu

Research output: Chapter in book / Conference proceedingConference article published in proceeding or bookAcademic researchpeer-review

Abstract

Since the usage of Unmanned Aerial Vehicle (UAV) for civil
applications is increasing, the localization accuracy in urban becomes an
important issue for safety. However, the GNSS localization solution suffers a
large error by the multipath effect. Since the multipath effect is unable to be
completely solved but to mitigate, the multi-sensor integrated localization
method is a common method to reduce this error. This study develops an adaptive Kalman filter adjusting the weighting between GNSS and INS measurements for different circumstance, further to improve the integration performance. The adaptation is based on supervised machine learning model classification, predicting the GNSS conditions with measurement features. The principle component analysis (PCA) is employed to aid selecting major features and labeling data for machine learning. Then, the supervised machine learning model is trained base on the decision tree and random forest (RF) learning algorithm with real operation data covering most situations. To reduce the missclassification error, the fuzzy logic algorithm is designed to avoid the
classification result with rapid change. Besides, the process noise covariance is
determined with Allan variance analysis. The localization performance of the
proposed adaptive Kalman filter is compared with conventional Kalman filter
and onboard localization solution provided by commercial fly controller. The
results prove that the presented adaptive Kalman filter using random forest with
fuzzy logic can achieve better GNSS condition classification outperforming other
algorithms. The fuzzy logic in the proposed algorithm can mitigate jumping error causing by miss-classification. For urban areas, the overall localization result improves about 50% comparing with the onboard solutions. The maximum
localization error can be reduced from 43.2 to 14.7 meters. The result verifies
that the proposed adaptive Kalman filter can mitigate the localization error from
multipath effect as well as achieving more accurate localization solution for UAV
in urban areas.
Original languageEnglish
Title of host publicationProceedings of International Symposium on GNSS, Hong Kong
Number of pages26
Publication statusPublished - 10 Dec 2017

Fingerprint

Dive into the research topics of 'Adaptive GNSS/INS Integration Based on Supervised Machine Learning Approach'. Together they form a unique fingerprint.

Cite this