Distributed estimation using online semi-supervised particle filter for mobile sensor networks

• Published in: IET control theory & applications 2015-02, Vol.9 (3), p.418-427
• Main Authors: Yoo, Jaehyun, Kim, Woojin, Kim, Hyoun Jin
• Format: Article
• Language: English
• Subjects: Accuracy; Algorithms; available labelled training data; communication bandwidth; Deviation; Distance learning; distributed estimation; Learning; learning (artificial intelligence); mobile sensor networks; MSN; Networks; obstacle configurations; online semisupervised particle filter; particle filtering (numerical methods); received signal strength indicator‐based distributed location estimation; semisupervised machine learning; Sensors; Signal strength; Special Issue on Co-operative Multi-Agent Systems with Engineering Applications; statistical analysis; statistical variability; suddenly‐changed RSSI characteristics; telecommunication computing; time‐varying prior model; wireless sensor networks
• ISSN: 1751-8644; 1751-8652; 1751-8652
• DOI: 10.1049/iet-cta.2014.0495

This study proposes an improved particle filter by incorporating semi-supervised machine learning for location estimation in mobile sensor networks (MSNs). A time-varying prior model is learned online as the likelihood of particle filter in order to adapt to dynamic characteristics of state and observation. Thanks to semi-supervised learning, the proposed particle filter can improve efficiency and accuracy, where the amount of available labelled training data is limited. The authors compare the proposed algorithm with the particle filter based on supervised learning. The algorithms are evaluated for received signal strength indicator (RSSI)-based distributed location estimation for MSN in which communication bandwidth and accuracy of the range measurement are limited. First, experimental results show that the semi-supervised algorithm can learn suddenly-changed RSSI characteristics while the supervised learning cannot. Second, the proposed particle filter is more accurate and robust against variations of the environment such as new obstacle configurations. Furthermore, the suggested particle filter shows low statistical variability during repeated experiments, confirmed by much smaller error deviation than the compared particle filter.