Highlighting Prominent Features for Size Reduction in Time Series Data using Clustering Techniques

Abstract

There exist many techniques for feature selection and reduction to reduce dimensions of the large sensor dataset. For real time data processing, compressed and prominent feature of highest significance is desirable for efficient way of resource optimization and computation cost reduction. The goal of this research study is to highlight most significant feature of the dataset and to generate compressed time series by highlighting it. The highlighted feature of accelerometer sensor dataset is extracted, and a more compressed form of time series is generated using statistical and clustering methods like k-means, Partition around Medoids (PAM), Max-Value, 95% Confidence Interval values and Ceil Function calculations. As a result, around 80 % reduction in dataset with the similar pattern as of original time series is achieved. The original time series is compared with generated output series using Dynamic Time Warping method, where, we have obtained normalized error distance of 0.02. (Accuracy 98%)