Early Classification Approaches for Sensors Generated Multivariate Time Series with Different Challenges

Early classification of Multivariate Time Series (MTS) that generated from sensors, has received a great attention as it has potential to solve time-critical problems of many areas including healthcare, industries, and intelligent transportation. A time series is also called as component if it is part or dimension of MTS. Unlike the existing work on early classification, this work considers following major challenges of MTS: 1) different length components, 2) faulty components, and 3) presence of unknown (unseen) class labels. We proposed different approaches for addressing these challenges of MTS in the framework of early classification. We also demonstrated the effectiveness of the approaches for the real-world applications such as road surface classification, human activity classification, and identification of faults in the washing machines. The experimental results showed that the proposed early classification approaches can achieve significant earliness with a marginal compromise of accuracy. Currently, we are working on the early classification of network applications such as Youtube, Firefox, and Skype, using the traffic flow of packets.