Clustering Time Series Sensor Data Using Modified Kohonen Maps

Abstract

With the increase in the usage of sensors to collect data, there has been a large increase in the number of time series data captured via these devices. These are of different varieties of them, ranging from astronomical to meteorological measurements. The ability to cluster these data allows us to not only process and prepare the data for further mining but also develop an important tool in compressing sensor data for better quality and faster communication. In this paper, we introduce a procedure using Kohonen Maps to cluster such data and compare it to the common procedure of Hierarchical clustering for times series instances. There are two modifications done to the conventional Kohonen Maps algorithm -1) The distance measure used is the DTW distance instead of the traditional Euclidean distance and 2) A sampling scheme is introduced which chooses the most diverse elements as the initial cluster representatives. The distance/similarity measure employed to compare them both is the dynamic time warping (DTW) measure, since there is enough literature to show its superior performance over other algorithms. The proposed algorithm was found to be better in terms of both quality of clusters obtained as well as speed when compared to Hierarchical clustering using DTW as a distance measure which is one of the most popular techniques of clustering time series data.