Robust data cleaning methodology using online support vector regression for ultra-short baseline positioning system.

Abstract

Ultrashort baseline (USBL) acoustic positioning system is a significant navigation means for human occupied vehicle due its simple structure, convenient operation, and large-scale-maneuver capacity. In order to improve the quality of USBL raw data effectively and efficiently, a robust data cleaning methodology using Online Support Vector Regression (OSVR) is proposed to deal with measurement outliers and missing values. In this study, we applied sliding-window samples to train the OSVR model for online time series prediction and then utilized the obtained one-step ahead prediction to detect and replace outliers or supplement missing values. The experimental results of the online test show that the proposed methodology can satisfy the requirement of real-time navigation and acquire consecutive and consistent positioning data for USBL. In comparison with the raw data, the root mean square error results in longitude and latitude are reduced by 91.75% and 85.53%, respectively. In addition, such methodology outperforms other data cleaning algorithms based on Least Square (LS) and kernel recursive LS.