Burst-Based Event Classification on Weakly Labeled Time Series Data of Sensors
Hanbo Zhang, Yawen Wang, Peng Wang, Wei Wang · 2017
Detecting events on time series data generated by sensors has received a great amount of attention with increasingly deployment of variable sensors. In this paper, we propose a novel framework for classifying events upon sensors data called BEC. Given long raw time series and event labels on fuzzy time points, BEC extracts burst-based features to represent the events. There are mainly two important tasks to be solved in our framework. First, we automatically extend fuzzy time points to appropriate subsequences containing sufficient information. Second, we extract burst-based features to train the classification model. We demonstrate on reallife datasets that without unrealistic assumptions and human interventions, our framework outperforms the state-of-the-art approaches when dealing with sensors data.